Bibliography in Assisted Reproductive Technology (ART)

In an ever-evolving field marked by rapid scientific progress and increasing clinical complexity, it is essential to remain updated on key innovations, best practices, and the most relevant publications.

Compiled and reviewed by the clinical team at the xxxxxx—one of Paris’ most established and active ART centers—this summary offers a concise yet rigorous overview of recent advances in the field. Drawing from high-impact international journals, the work highlights emerging techniques, evolving protocols, and critical findings in ART.

With this bibliographical review, we aim to offer an original tool to aid in scientific monitoring, positioned between the conventional abstract and the often lengthy and time-consuming full reading of the original article. This hybrid format ensures substantial time savings by enabling the rapid and targeted identification of publications that, based on their content, warrant an in-depth reading.

Regular updates will follow to keep pace with the dynamic landscape of ART.

Artificial Intelligence in Assisted Reproductive Technology: Enhancing Precision, Efficiency, and Outcomes in Modern Fertility Care

Dr Cécile François

Hm-envelop Linkedin

Dr François durand

Hm-envelop Youtube Linkedin

Overview

The landscape of modern society, characterized by career-centric lifestyles and the global trend of postponing childbirth, has led to a significant increase in infertility. Globally, an estimated one in six couples grapples with infertility, a prevalence that can surge to 30% in some regions. Despite significant technological advancements in Assisted Reproductive Technology (ART) since the birth of the first “test-tube baby” in 1978, success rates, such as the live birth rate, typically hover around 30% per IVF cycle. This indicates substantial room for improvement and highlights persistent challenges, particularly given the labor- and time-intensive nature of IVF and the variability in outcomes due to inter- and intra-observer differences.
Artificial Intelligence (AI), a sophisticated technological framework emulating human cognitive functions, is increasingly recognized as an indispensable tool in healthcare. Its application in medicine spans from improving medical imaging interpretation and assisting in genomic analysis to aiding precise clinical decision-making, accelerating drug discovery, easing patient monitoring, and enhancing surgical procedures through robotic assistance. Within ART, AI, by leveraging the benefits of automation, promises to augment efficiency, reproducibility, and consistency, ultimately enhancing both treatment effectiveness and clinical decision-making. This synthesis will detail how AI is poised to revolutionize modern fertility care by enhancing precision, efficiency, and outcomes across various stages of ART.

At its core, AI endows computational systems with the capacity for learning, comprehension, problem-solving, and task execution, akin to human reasoning. Machine learning (ML), a crucial component of AI, involves algorithms designed to learn from data and make predictions or decisions.

  • Types of Machine Learning Algorithms and their ART Applications:
    • Supervised Learning: This approach trains algorithms using labeled datasets where each data point is associated with a known outcome. In ART, it is frequently employed for tasks such as embryo classification, sperm analysis, and the prediction of treatment outcomes. Its accuracy is significantly influenced by the quality and quantity of training data.
    • Unsupervised Learning (UL): Unlike supervised learning, UL autonomously processes extensive, unlabeled datasets to uncover hidden patterns and structures. This independence is vital for improving embryo selection, tailoring patient treatments, identifying new factors affecting ART outcomes, and enhancing treatment protocols.
    • Reinforcement Learning (RL): RL involves an AI system learning by dynamically interacting with its environment through a trial-and-error method, balancing exploration and exploitation. In ART, RL can significantly improve treatment protocols, particularly in hormone therapy for IVF, by analyzing patient responses to personalize treatment plans.
    • Artificial Neural Networks (ANN) and Deep Learning: ANNs, inspired by biological neural networks, employ nodes (neurons) communicating through connections with varying weights. Deep learning, an evolution of classic ML, utilizes deep neural networks (DNNs) that play a pivotal role in recognizing patterns in complex medical data, including medical images. Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) are specific deep learning methods applied for tasks like oocyte image analysis and sperm quality assessment.
  • Data Input and Processing: Effective AI training in ART necessitates extensive, high-quality data, including patient demographics, treatment protocols, and medical histories. Advanced imaging, like time-lapse embryo monitoring, provides critical insights into embryo development. However, obtaining large, diverse, and high-quality datasets for AI models can be challenging due to the sensitive and variable nature of reproductive data, and issues with data standardization, collection, and sharing persist.

AI is being integrated across various stages of the IVF process to enhance clinical decision-making and optimize treatment efficiency.

  • Controlled Ovarian Stimulation (COS): COS is a critical component of IVF aimed at harvesting an ideal quantity of mature Metaphase II (MII) oocytes while mitigating the risk of Ovarian Hyperstimulation Syndrome (OHSS).
    • Optimizing FSH Dosage: AI is playing a pivotal role in automating stimulation progression, calibrating dosage adjustments, and precisely timing interventions. Machine learning models, leveraging historical clinical data, are being developed to personalize FSH doses for IVF, aiming to maximize mature oocyte yield and minimize OHSS risk. Studies have shown AI models can outperform clinicians in precisely targeting the retrieval of 10-15 eggs and personalize dose-response profiles, potentially leading to more mature oocytes and reduced gonadotropin consumption.
    • Follicle Monitoring and Trigger Timing: AI-aided ultrasound, including 3D ultrasound, enhances the evaluation of follicle sizes and numbers, and may even enable at-home fertility assessments. AI models can optimize the day of trigger to improve IVF outcomes, effectively increasing the number of fertilized oocytes and blastocysts by leveraging parameters such as follicle size and estradiol levels.
  • Protocol Selection: AI research is dedicated to refining IVF protocol strategies, with meta-analyses advocating for GnRH antagonist protocols due to their reduced OHSS risk. AI evaluations affirm that antagonist protocols are suitable for poor responders and that the specific protocol chosen has minimal effect on success rates.
  • Workflow Optimization: Algorithms can identify optimal monitoring days, forecast trigger timing, and estimate oocyte yields, reducing patient visits and optimizing lab scheduling, thus improving the efficiency and predictability of IVF workflows.
  • Patient Preparation: AI-driven tools are emerging to assist in patient preparation, such as predicting implantation by evaluating endometrial thickness and receptivity with greater accuracy than traditional methods.
  • Clinical Decision Support Systems (CDSS): Beyond specific IVF steps, AI is used to develop CDSSs to assist in clinical decision-making in reproductive medicine. For instance, AI can predict patients who would improve sperm parameters after varicocele repair or identify non-obstructive azoospermia (NOA) patients who may benefit from sperm extraction attempts.

AI significantly enhances the precision of selecting embryos and sperm, and advances the prediction of fertilization and genetic outcomes, providing embryologists and clinicians with sophisticated tools to improve pregnancy prospects.

  • Oocyte Morphological Assessments: AI and Deep Learning (DL) technologies are transforming oocyte quality assessment by offering objective and consistent analysis of meiotic stages (MII, MI, GV), enabling a more accurate prediction of fertilization potential.
    • DNN-based systems can classify human oocytes by meiotic maturity with high accuracy (96.4% in validation, 95.7% in testing).
    • Enhanced U-Net models precisely segment key characteristics like ooplasm, zona pellucida, and perivitelline space in low-resolution MII oocyte images, showing superior accuracy over traditional methods.
    • CNNs and SVMs have improved the prediction of oocyte developmental potential, with CNN models capable of predicting fertilization potential with over 86% accuracy and distinguishing mature MII oocytes with 98.9% precision by pinpointing the first extruded polar body. This precision aids in identifying optimal sites for sperm injection during ICSI.
    • The aim is to identify specific oocyte features (physical traits, cellular formations, morphometric and morphokinetic parameters, cytoplasmic movements) that embryologists use, to leverage them for AI applications, focusing on live birth potential. AI can unlock hidden patterns in collected clinical data and improve decision-making.
  • Semen Analysis: AI and DL are transforming sperm selection for IVF and ICSI by expanding beyond traditional semen parameter analysis to automated sperm detection, comprehensive semen analysis, and assessments of sperm viability and DNA integrity.
    • AI systems, trained on detailed sperm datasets, accurately identify optimal sperm morphology and classify sperm by motility, improving upon the variability of traditional methods. Automated classifiers can achieve up to 94% accuracy rates in classifying sperm morphology.
    • Deep learning algorithms can accurately detect morphological abnormalities in sperm acrosome and vacuole regions in real-time on standard laptops.
    • Computer-Assisted Sperm Analysis (CASA) systems, enhanced with AI, allow for precise and quick analysis of sperm trajectories, improving accuracy and efficiency in semen analysis and male infertility prediction.
    • AI also shows promise in predicting sperm with high DNA fragmentation rates, a known cause of male infertility and ART failure.
  • Embryo Selection and Ploidy Prediction: AI-based tools offer a standardized, non-invasive approach to prioritizing high-quality embryos for transfer by analyzing morphological features and predicting euploidy (normal chromosome number).
    • Time-lapse incubators, combined with AI, provide continuous monitoring, aiding in selection.
    • AI models show potential in accurately predicting embryo ploidy from images and videos, with accuracy ranging from 44% to 85%, and can be improved by combining image analysis with clinical data. For example, one AI model effectively predicts embryo ploidy with 65.3% accuracy, improving to 77.4% post-optimization, and showing high-scoring embryos by AI to be twice as likely to be euploid.
    • Studies demonstrate AI’s potential to significantly decrease the number of cycles required to reach a clinical pregnancy and increase the first cycle clinical pregnancy rate when used as an adjunct to embryologists’ expertise. The EMBRYOLY algorithm, for example, could have increased the first cycle pregnancy rate from 19.8% to 44.1% and reduced cycles to clinical pregnancy from 2.01 to 1.66.
    • This is particularly beneficial for clinics where PGT-A is not common or available, as AI can assist in ranking embryos without prior knowledge of ploidy status. AI can provide reproducible recommendations and enhance objectivity in embryo assessment, harmonizing practices across centers and countries.
    • AI can also predict the developmental potential of cleavage-stage embryos, distinguishing between good and poor quality, and assisting in sorting embryos with high quality when embryologists face difficulty in selection.
  • Micromanipulation Procedures: AI, with its advanced image-processing capabilities, is revolutionizing embryology by providing precise guidance for intricate procedures such as ICSI and assisted hatching.
    • Deep learning CNNs can identify critical morphological features in oocytes and embryos, enhancing precision in ICSI and assisted hatching (AH) procedures with accuracies of 98.9% and 99.41%, respectively.
    • Automated Intracytoplasmic Sperm Injection (ICSIA) robots have demonstrated effectiveness by automating key IVF steps, achieving high survival rates for injected human oocytes and leading to successful fertilizations and births.
    • AI can detect the optimal location for ICSI to avoid membrane ruptures, offering a promising non-invasive tool to increase IVF outcomes.

AI’s application extends to refining predictive modeling, enhancing patient safety, and ensuring quality management within ART.

  • Predicting Clinical Pregnancy and Live Birth: Machine learning refines predictive modeling in IVF by leveraging patient-specific data to enhance the accuracy of outcomes such as live birth and clinical pregnancy rates.
    • Random forest models have been found more effective than logistic regression in predicting clinical pregnancies, identifying factors like ovarian stimulation protocols as positive predictors and female age and infertility duration as negative influences.
    • Hybrid AI models integrating time-lapse video data with clinical features (e.g., embryo morphokinetics, oocyte age, gonadotropin dosage, endometrium thickness) have shown significant improvement in predicting clinical pregnancy outcomes.
    • Adaptive data-driven models can accurately predict live birth outcomes, enabling real-time updates and decision support throughout the IVF process.
  • Social Egg Freezing: AI models provide precise age-related guidelines for fertility preservation, helping determine the optimal number of oocytes to freeze for a desired chance of live birth. For instance, women under 35 may need to freeze 15 eggs for a 70% chance of live birth, while at 42, 61 eggs might be required for a 75% success rate.
  • Patient Safety: The critical nature of IVF lab procedures demands flawless execution to prevent human error, such as gamete loss or transfer of mismatched gametes.
    • AI, specifically CNNs, can precisely identify IVF embryo stages from time-lapse images, achieving 100% accuracy in identifying patients from their embryos, enhancing tracking accuracy in IVF laboratories.
    • Electronic witnessing systems (EWS), utilizing RFID and barcodes, have effectively prevented sample mix-ups in assisted reproduction over a decade, achieving very low mismatch rates and critical errors. These systems are crucial for tracking gametes and embryos from collection to cryopreservation and transfer, reducing human error.
  • Quality Management: AI contributes to standardizing procedures and improving reproducibility, addressing issues like fragmentation and variability across IVF laboratories. By automating routine and time-consuming steps, AI not only guarantees optimal practices and results but also reduces the likelihood of errors.

Despite AI’s transformative potential, its widespread implementation in IVF faces significant hurdles.

  • Implementation Challenges and Slow Adoption: Practical challenges include high implementation costs for hardware and software, the need for extensive staff training, and complexities in integrating AI into existing clinical workflows. Despite projected global AI market growth, its penetration in IVF remains modest compared to other medical sectors, reflecting a broader truth that new technologies require demonstrable, significant advantages to achieve rapid clinical integration.
  • Data Concerns: AI models rely heavily on large, diverse, and high-quality datasets for training, which can be difficult to obtain given the sensitive and variable nature of reproductive data. Lack of standardization in image capture and data formats across clinics can introduce bias and limit the generalizability and robustness of AI models. Many studies have been tested on limited populations, raising questions about their generalizability.
  • Ethical and Regulatory Hurdles: Significant ethical and regulatory concerns include data privacy (e.g., HIPAA, GDPR), legal liability due to the opacity of AI systems (“black boxes”), potential erosion of clinician skills due to over-reliance, and societal pushback on genetic predictions (e.g., “designer babies”). Regulatory frameworks often lag behind AI’s rapid evolution, and global harmonization remains limited, although efforts like the EU AI Act and Good Machine Learning Practices are emerging. Transparency and explainability in AI systems are critical to inspire trust and avoid misinterpretation or blind adoption of outputs.
  • Sustainability: The increasing energy consumption of data centers supporting AI raises environmental sustainability concerns, with projections showing a significant rise in electrical consumption.
  • Future Directions: The future of AI in ART is poised for further innovation:
    • Personalized Treatment: AI can further tailor IVF treatments by optimizing drug dosing, timing for antagonists, reducing OHSS risk, and perfecting trigger methods.
    • Advanced Diagnostics: Integrating multi-omics data with AI can deepen the understanding of genetic and biochemical factors influencing embryo viability, ovarian response, and implantation.
    • Automation Expansion: Automation is expanding into sperm preparation, oocyte retrieval and denudation, dish preparation, and cryopreservation, with new systems under evaluation. AI could also simplify and automate microfluidics for autonomous IVF processes, reducing reliance on embryologist tasks.
    • Earlier Oocyte Assessment: There is a compelling need to focus analysis on pre-ICSI oocytes, aiming to anticipate analysis and selection of oocytes before fertilization to minimize embryo wastage and optimize IVF outcomes, potentially shortening time to pregnancy and reducing the number of frozen embryos.
    • Patient Empowerment: AI can enhance patient-clinician communication and trust by providing precise, data-driven recommendations, enabling patients to feel more confident in their treatment choices.
  • Collaborative Approach: The integration of AI into ART necessitates a balanced approach, recognizing AI as an ally that complements clinicians’ expertise, rather than a competitor. Success hinges on embryologists and clinicians embracing AI, sharing high-quality data, and collaborating in refining these tools. This collaborative commitment will redefine patient outcomes and ensure ART remains at the forefront of innovation and care excellence. The physician’s task in the future will be to work in an integrated manner with machine learning without blindly relying on it.

In conclusion, AI represents a transformative advancement for ART, promising to significantly improve the precision, efficiency, and outcomes of fertility treatments. While challenges related to implementation, data, ethics, and regulation persist, a collaborative, rigorous, and patient-centric approach to AI development and integration will be key to unlocking its full potential in reproductive medicine.

FAQ

  • Supervised learning: Algorithms are trained using labeled datasets to make future predictions, frequently employed for tasks such as embryo classification, sperm analysis, and the prediction of treatment outcomes.
  • Unsupervised Learning (UL): Processes extensive datasets autonomously without needing human direction to uncover hidden patterns and structures, proving valuable for improving embryo selection, tailoring patient treatments, identifying new factors affecting ART outcomes, and enhancing treatment protocols.
  • Reinforcement Learning (RL): An AI system learns by dynamically interacting with its environment, balancing exploration and exploitation to achieve set goals, and is especially effective in optimizing hormone therapy in IVF by analyzing patient responses. Other classic machine learning models include linear regression, logistic regression, decision trees, random forests, and Support Vector Machines (SVM). Artificial Neural Networks (ANN) are also an evolution of classic machine learning, inspired by biological neural networks.

AI plays a pivotal role in automating stimulation progression, calibrating dosage adjustments, and ensuring the precision of trigger timing to maximize the yield of mature Metaphase II (MII) oocytes. It also contributes to the efficiency of follicle monitoring. Machine learning models, utilizing historical clinical data, streamline the selection of initial Follicle-Stimulating Hormone (FSH) doses for IVF, providing a standardized framework to enhance the personalization of treatment protocols and mitigate variability in ovarian responses. Studies have shown that AI models can be superior to clinicians in more precisely targeting the retrieval of 10-15 eggs.

AI-aided ultrasound is transforming follicular monitoring by facilitating automated monitoring capabilities, with the potential to enable at-home fertility assessments comparable to clinic visits. It helps in accurately measuring leading follicles, which are tracked until they reach 16-22 mm, the optimal size for HCG trigger administration. AI models, such as CR-Unet and HaTU-Net, specifically enhance transvaginal ultrasound for precise segmentation of ovarian and follicular structures, which is crucial for fertility treatments, improving clinical diagnostics and treatment planning. 3D ultrasound coupled with AI integration offers a comprehensive volumetric view that surpasses the limitations and operator variability of traditional 2D ultrasound.

AI integrates a wide array of data to enhance the precision of determining the optimal moment for the trigger injection, which is crucial for maturing oocytes. A machine learning causal inference model (a T-learner with bagged decision trees), applied to a large number of IVF patients, demonstrated its capability to enhance the timing of trigger injections, surpassing physician judgment. This model effectively increased the number of fertilized oocytes and blastocysts by leveraging key parameters such as follicle size and estradiol levels. Advanced algorithms can also predict the optimal monitoring day and forecast trigger timing within a three-day window, which reduces patient visits and optimizes lab scheduling, improving the efficiency and predictability of IVF workflows.

AI and Deep Learning (DL) technologies are transforming oocyte quality assessment in IVF labs by offering objective and consistent analysis of meiotic stages (Metaphase II, Metaphase I, and Germinal Vesicle) non-invasively. This enables a more accurate prediction of fertilization potential, leading to improved pregnancy outcomes. A DNN-based system has been introduced that classifies human oocytes by meiotic maturity with high accuracy, achieving 96.4% in validation and 95.7% in testing. Enhanced U-Net models precisely segment key characteristics like ooplasm, zona pellucida, and perivitelline space from low-resolution images, improving automated assessment of oocyte quality and embryo implantation potential. Additionally, Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) have markedly improved the prediction of oocyte developmental potential, refining IVF protocols and ICSI procedures.

AI and DL technologies significantly enhance sperm selection for IVF and ICSI by automating sperm detection, providing comprehensive semen analysis, and assessing sperm viability and DNA integrity. These AI systems accurately identify optimal sperm morphology and classify sperm by motility, improving upon the variability seen in traditional methods. A novel deep learning algorithm accurately detects morphological abnormalities in sperm acrosome and vacuole regions in real-time on standard laptops. AI, specifically through CNNs, allows for precise and quick analysis of sperm trajectories, greatly improving the accuracy of semen analysis and predicting motility from video data alone. AI models can also predict sperm with a high DNA fragmentation rate and assess DNA integrity from brightfield images, aiding in the selection of high-integrity sperm for ICSI procedures. Computer-aided sperm analyzers (CASA), augmented by AI, are useful tools for rapid sample analysis, reducing interoperator variability and providing accurate measures of sperm motility and kinematics.

AI models show potential in accurately predicting embryo ploidy from images and videos, which may enhance IVF outcomes by enabling better embryo selection. A new AI model, trained on Day 5 blastocyst images from various IVF clinics, effectively predicts embryo ploidy, with higher-scored embryos by the AI being twice as likely to be euploid. AI systems like ERICA predict ploidy potential and probability of pregnancy from day 5 embryo images, potentially surpassing human embryologists in accuracy. The integration of morphological and clinical data enhances embryo ploidy prediction compared to traditional preimplantation genetic testing for aneuploidy (PGT-A), offering more precise and non-invasive embryo selection for improved IVF outcomes.

Traditional morphological scoring for embryo selection can be subjective and inconsistent, leading to arbitrary selections, especially when many high-quality embryos are available. AI-based tools, conversely, offer a standardized, non-invasive approach to prioritizing high-quality embryos for transfer, analyzing morphological features to estimate the likelihood of pregnancy. While AI reduces assessment time and standardizes evaluations, current evidence does not conclusively show universally better clinical outcomes compared with traditional methods, indicating a need for more validation through large prospective trials. However, studies suggest that AI, when used adjunctively with embryologist expertise, can statistically decrease the number of cycles required to reach a clinical pregnancy and increase the first cycle clinical pregnancy rate by better ranking sibling embryos.

Machine learning refines predictive modeling in IVF by leveraging patient-specific data to enhance the accuracy of clinical outcome predictions, such as live birth and clinical pregnancy rates. Models like Random Forest have shown higher accuracy in predicting clinical pregnancies compared to logistic regression, emphasizing the impact of ovarian stimulation protocols and identifying factors such as female age and infertility duration as negatively influencing outcomes. Hybrid AI models integrating time-lapse video data with clinical features (e.g., embryo morphokinetics, oocyte age, gonadotrophin dosage, and endometrium thickness) significantly enhance the prediction of clinical pregnancy outcomes. Furthermore, large multi-center studies utilizing dynamic, data-driven models with various factors can accurately predict live birth outcomes, enabling real-time updates and decision support throughout the IVF process.

The critical nature of IVF lab procedures demands flawless execution to prevent severe repercussions of human error, such as gamete loss or the transfer of mismatched gametes. AI, specifically through CNNs, has been shown to precisely identify IVF embryo stages from time-lapse images, achieving 100% accuracy in identifying patients based on their embryos on days 3 and 5, thereby enhancing tracking accuracy and avoiding human error. Electronic witnessing systems (EWS), often using RFID and barcodes, effectively prevent sample mix-ups, achieving very low mismatch rates and proving especially reliable in sperm preparation and IVF/ICSI procedures. These systems are crucial for maintaining the integrity and safety of fertility treatments.

AI, with its advanced image-processing capabilities, is revolutionizing embryology by providing precise guidance for intricate procedures such as Intracytoplasmic Sperm Injection (ICSI) and assisted hatching (AH). Deep learning CNNs have been utilized to identify critical morphological features on oocytes and embryos, enhancing precision in ICSI and AH procedures with high accuracies (98.9% and 99.41% respectively), moving towards automating these intricate tasks. In a groundbreaking study, an Automated Intracytoplasmic Sperm Injection (ICSIA) robot demonstrated its effectiveness by automating key IVF steps, with survival rates for injected human oocytes close to manual methods, and has even led to successful births.

AI models provide precise age-related fertility preservation guidelines by integrating clinical values such as age, Antral Follicle Count (AFC) by ultrasound, and Anti-Müllerian Hormone (AMH) levels. For example, AI suggests that women under 35 should freeze 15 oocytes for a 70% chance of a live birth, and 25 oocytes for a 95% chance. For women at 34, freezing 10 oocytes offers a 75% chance; at 37, 20 oocytes are needed for the same outcome, and at 42, 61 oocytes are necessary for a 75% success rate. These insights support personalized fertility strategies, enhancing patient-doctor discussions and boosting confidence in medical choices.

  • Enhanced Precision and Accuracy: AI refines tasks like embryo and sperm analysis, provides more precise forecasts for clinical interventions, and improves diagnostic and predictive accuracy.
  • Increased Efficiency and Reproducibility: Automation of routine tasks alleviates workload for physicians and embryologists, standardizes practices, and reduces inter- and intra-observer variability.
  • Personalized Treatment Plans: AI can tailor treatment plans (e.g., hormone therapy, FSH dosage, protocol selection) based on individual patient responses and historical data.
  • Improved Outcomes: Leads to higher success rates in fertility treatments, including improved live birth and clinical pregnancy rates.
  • Reduced Human Error: AI helps avoid errors in identifying embryos and standardizes assessments, reducing subjectivity.
  • Accessibility and Cost-Effectiveness: Automation may make fertility care more affordable and accessible, and AI can reduce medical costs. Portable ultrasound systems enable at-home fertility assessments.
  • Enhanced Communication and Trust: Providing precise, data-driven recommendations helps patients feel more confident in their treatment choices and the guidance provided by their healthcare providers.
  • High Implementation Costs: Substantial investment in technology, hardware, and personnel development is required.
  • Data Quality and Availability: AI models depend on large, diverse, high-quality, and annotated datasets, which are difficult to obtain due to the sensitive and variable nature of reproductive data, lack of standardization, and privacy concerns.
  • Regulatory Hurdles and Ethical Concerns: Navigating complex regulations, ensuring patient privacy, addressing biases in models, and concerns about depersonalization of treatment or “designer babies” are significant challenges.
  • Integration Challenges: Ensuring system compatibility and seamless integration into existing clinical workflows with diverse IVF protocols and equipment can be difficult.
  • “Black Box” Problem: Many AI systems operate as “black boxes,” lacking transparency, which could erode trust and lead to misinterpretations or blind adoption.
  • Erosion of Human Expertise: Over-reliance on AI may degrade clinicians’ and embryologists’ decision-making skills.
  • Sustainability Concerns: Energy-intensive computational processes and expanding data centers raise environmental issues.
  • Lack of Clear Clinical Applications: Much of the current literature tends to focus on technical aspects without clear emphasis on outcome-driven differences or tangible clinical applications, limiting their relevance to real-world IVF practices.

Data is essential for training AI models that recognize patterns and improve predictions over time. Effective AI training in ART requires extensive data on patient demographics, treatment protocols, and medical histories. Data quality, standardization, and extracting relevant information from academic sources are crucial for refining AI models and enhancing their accuracy. Without diverse and high-quality datasets, AI models can lack robustness, struggle to generalize to real-world scenarios, and potentially perpetuate disparities if trained on non-representative data. Multi-center studies and uniformity in data generation procedures are needed to reduce biases and increase the overall robustness of generated AI-based models.

Bibliography

  • Wu, Y.-C., Su, E. C.-Y., Hou, J.-H., Lin, C.-J., Lin, K. B., & Chen, C.-H. (2025). Artificial intelligence and assisted reproductive technology: A comprehensive systematic review. Taiwanese Journal of Obstetrics & Gynecology, 64, 11–26.
  • Calogero, A. E., Crafa, A., Cannarella, R., Saleh, R., Shah, R., & Agarwal, A. (2024). Artificial intelligence in andrology – fact or fiction: essential takeaway for busy clinicians. Asian Journal of Andrology, 26, 600–604.
  • Cohen, J., Silvestri, G., Paredes, O., Martin-Alcala, H. E., Chavez-Badiola, A., Alikani, M., & Palmer, G. A. (2025). Artificial intelligence in assisted reproductive technology: separating the dream from reality. Reproductive BioMedicine Online.
  • Dissler, N., Nogueira, D., Keppi, B., Sanguinet, P., Ozanon, C., Geoffroy-Siraudin, C., Pollet-Villard, X., & Boussommier-Calleja, A. (2024). Artificial intelligence-powered assisted ranking of sibling embryos to increase first cycle pregnancy rate. Reproductive BioMedicine Online.
  • Wang, X., Wei, Q., Huang, W., Yin, L., & Ma, T. (2024). Can time-lapse culture combined with artificial intelligence improve ongoing pregnancy rates in fresh transfer cycles of single cleavage stage embryos? Frontiers in Endocrinology, 15, 1449035.
  • Iannone, A., Carfì, A., Mastrogiovanni, F., Zaccaria, R., & Manna, C. (2024). On the role of artificial intelligence in analysing oocytes during in vitro fertilisation procedures. Artificial Intelligence in Medicine, 157, 102997.

The study analyzed 48 relevant articles to assess the impact of AI on aspects such as treatment efficacy, process optimization, and outcome prediction. A particular emphasis was placed on examining the effectiveness of different machine learning paradigms—supervised, unsupervised, and reinforcement learning—in improving ART-related procedures.

  1. Context and the Challenge of Infertility

In the contemporary era, a global trend of postponing childbirth has led to infertility becoming a significant concern. It is estimated that one in six couples worldwide contends with the challenges of infertility. In the United States, approximately 8.8% of the population is affected, while in regions beyond the US, this figure ranges from 8% to 12%, with some areas experiencing a prevalence as high as 30%. In Taiwan, where birth rates are among the lowest globally, 10-15% of couples grapple with infertility, and the average duration for diagnosing infertility stretches to 2.9 years, significantly exceeding the WHO’s one-year benchmark. Treatment typically commences a further 1.5 years later. Data from 2021 indicates that in Taiwan, 55.2% of ART treatment cycles were due to female factors, 8.3% to male factors, 32.2% to combined factors, and 4.3% remained unexplained.

ART encompasses fertility treatments that involve handling eggs or embryos, including surgical egg removal, laboratory fertilization, and transfer back to a woman’s body or to another woman. This definition specifically excludes treatments like intrauterine insemination (IUI) and timed intercourse. Despite significant technological advancements in IVF, success rates remain challenging. The implantation rate in cycles without preimplantation genetic testing for aneuploidy (PGT-A) is only close to 50%, and even with PGT-A, it increases to around 60% at its highest. The live birth rate typically hovers around 30% in each IVF cycle. IVF is a sophisticated, multistage procedure that is labor- and time-intensive, coupled with significant inter- and intra-observer variability, which impacts its reproducibility and efficiency.

  1. The Role and Benefits of AI in ART

AI is increasingly being employed across various medical applications, enhancing medical imaging interpretation, assisting in genomic analysis for personalized treatments, and playing a role in precise clinical decision-making. It also accelerates drug discovery, eases patient monitoring, and enhances surgical procedures through robotic assistance. Within ART, AI, by leveraging automation, assumes a promising role in augmenting efficiency, reproducibility, and consistency.

AI has the potential to alleviate the workload of physicians and embryologists by automating routine and time-consuming steps in IVF, such as ovarian stimulation or workflows in the embryology laboratory. Furthermore, through continuous machine learning to enhance accuracy, AI not only guarantees optimal practices and results but also reduces the likelihood of errors. The review underscores that AI technologies significantly enhance ART processes by refining tasks such as embryo and sperm analysis and facilitating personalized treatment plans based on predictive modeling. Notable improvements have been observed in the accuracy of diagnosing and predicting successful outcomes in fertility treatments. AI-driven models provide more precise forecasts of the optimal timing for clinical interventions such as egg retrieval and embryo transfer, which are critical to the success of ART cycles. The integration of AI into ART represents a transformative advancement, substantially improving the precision and efficiency of fertility treatments. AI is becoming an indispensable tool in reproductive medicine, enhancing both the effectiveness of treatments and the clinical decision-making process.

  1. Review Methodology

The review conducted a systematic literature search using PubMed, EMBASE, and Cochrane Library databases. The search included keywords like “artificial intelligence and infertility,” “artificial intelligence and in vitro fertilization (IVF),” and “artificial intelligence and assisted reproductive technology.” Additionally, cross-referenced terms such as AI, IVF, ART, ovarian stimulation, semen analysis, embryo transfer, endometrial preparation, follicle, ultrasound, machine learning, and deep learning were used. The comprehensive literature review was undertaken up to March 2024, focusing on research from the past five years. This search initially identified 5791 publications, which were meticulously screened for relevance, and duplicates were eliminated, resulting in 35 articles for in-depth review.

  1. Fundamental Principles of AI Technology

At its essence, AI is a sophisticated technological framework that emulates human cognitive functions, endowing computational systems with the capacity for learning, comprehension, problem-solving, and task execution.

  • Data Input and Processing: In AI, data is essential for training models that recognize patterns and improve predictions over time. Effective AI training in ART requires extensive data on patient demographics, treatment protocols, and medical histories. Advanced imaging, such as time-lapse embryo monitoring, provides critical insights into embryo development, aiding in pattern recognition and decision-making. Data quality, standardization, and extracting relevant information from academic sources are crucial for refining AI models and enhancing their accuracy in ART.
  • Machine Learning Algorithms: These algorithms enable computers to learn from and make predictions or decisions based on data. They encompass a wide range of techniques:
    • Supervised Learning: Involves training algorithms using labeled datasets, where each data point is associated with a label or target variable. This allows the algorithm to make predictions or classify data accurately. In ART, this approach is frequently employed for tasks such as embryo classification, sperm analysis, and the prediction of treatment outcomes.
    • Unsupervised Learning (UL): Processes extensive datasets without needing human direction, making it crucial in neural networks and deep learning applications. This includes clustering, anomaly detection, and dimensionality reduction. In ART, UL proves invaluable for improving embryo selection, tailoring patient treatments, identifying new factors affecting ART outcomes, and enhancing treatment protocols.
    • Reinforcement Learning (RL): Stands out for its role in sequential decision-making. An AI system learns by dynamically interacting with its environment, using a trial-and-error method, earning rewards for correct actions and penalties for mistakes. In ART, RL can significantly improve treatment protocols, particularly in hormone therapy for IVF, by analyzing patient responses to tailor treatment plans more personally, potentially increasing success rates.

The choice of machine learning approach (supervised, unsupervised, or reinforcement learning) in developing AI applications for ART should be driven by the specific task, the nature of the data, and available resources, as each method offers distinct advantages.

  1. AI Applications in IVF

AI provides comprehensive assistance in various ART treatments and at different stages of the IVF process, including ovarian stimulation, diverse aspects of the IVF laboratory, embryo transfer, and patient safety and quality management.

  • Controlled Ovarian Stimulation (COS): This is an integral part of IVF aimed at harvesting an ideal quantity of mature Metaphase II (MII) oocytes while mitigating the risk of Ovarian Hyperstimulation Syndrome (OHSS). AI plays a pivotal role in the automation of stimulation progression, the calibration of dosage adjustments, the precision of trigger timing to maximize MII oocyte yield, and the efficiency of follicle monitoring.
    • Initial Gonadotrophin Dose and Further Dose Adjustment: Proper calibration of the initial FSH dose (ranging from 150 to 450 IU) is vital for obtaining sufficient mature oocytes and mitigating OHSS risk. Machine learning models are being advanced to streamline the selection of FSH doses.
      • Correa et al. (2024) evaluated a machine learning model that proved superior to clinicians in precisely targeting the retrieval of 10–15 eggs, with development and validation scores of 0.87 and 0.89, respectively.
      • Fanton et al. (2022) employed a K-nearest neighbors model that personalized dose-response profiles, forecasting the collection of an additional 1.5 mature MIIs on average for dose-responsive patients. For others, selecting a lower dose could save around 1375 IUs of FSH without adverse effects.
      • Letterie et al. (2020) assessed an algorithm that showed high decision-making accuracy (0.92 for continuing/stopping treatment, 0.96 for trigger/cancel, and 0.82 for medication adjustment).
      • The IDoser machine learning model is currently being evaluated for its accuracy in determining optimal initial FSH doses in a multicenter trial.
    • Protocol Selection: AI endeavors to refine protocol selection strategies, aiming to bolster the precision and efficacy of predicting IVF outcomes. Murillo et al. (2023) affirm that antagonist protocols are suitable for poor responders and that the specific protocol chosen has minimal effect on success rates.
    • AI in Ultrasound Analysis: Transvaginal ultrasound (TVUS) is essential for tracking follicular growth, but manual measurement is time-intensive and subjective. AI facilitates automated monitoring, potentially enabling at-home fertility assessments. Studies indicate that follicles measuring 12-19 mm on the day of HCG trigger are most likely to yield a mature oocyte, while 19-24.5 mm tend to produce high-quality blastocysts.
      • Liang et al. (2022) pinpointed follicle volume thresholds using a deep learning segmentation model from 3D ultrasound data that more accurately predict oocyte maturity and optimize HCG timing.
      • CR-Unet and HaTU-Net, adaptations of U-Net, enhance transvaginal ultrasound for follicular tracking, offering precise segmentation.
  • AI and Decision-Making for Best Day of Trigger: Determining the optimal moment for the trigger injection is a nuanced decision.
    • Hariton et al. (2021) causal inference model (T-learner with bagged decision trees) for 7866 IVF patients demonstrated its capability to enhance the timing of trigger injections, surpassing physician judgment, effectively increasing the number of fertilized oocytes and blastocysts.
    • An advanced algorithm identified an optimal monitoring day, forecasts trigger timing within three days, and estimates oocyte yields, reducing patient visits and optimizing lab scheduling.
    • Robertson et al. (2021) found that scheduling follicular tracking scans on Days 5-7 of stimulation is key for optimal IVF planning.
  • AI and Laboratory Procedures: AI enhances the precision of selecting embryos and sperm and advances the prediction of fertilization and genetic outcomes.
    • Oocyte Morphological Assessments: AI and Deep Learning (DL) transform oocyte quality assessment, offering objective and consistent analysis of meiotic stages (MII, MI, GV) to predict fertilization potential.
      • Targosz et al. (2023) introduced a DNN-based system that classifies human oocytes by meiotic maturity with high accuracy (96.4% in validation and 95.7% in testing).
      • Firuzinia et al. (2021) developed an enhanced U-Net model for segmenting MII oocytes in low-resolution images, showing superior accuracy over traditional methods.
      • CNNs and SVMs have notably improved the prediction of oocyte developmental potential. A CNN model achieved over 86% accuracy in predicting oocyte fertilization potential.
    • AI and Semen Analysis: AI/DL technologies are transforming sperm selection for IVF and ICSI by automating sperm detection, providing comprehensive semen analysis, and assessing sperm viability and DNA integrity.
      • Javadi et al. (2019) developed a deep learning algorithm that accurately identifies morphological abnormalities in sperm acrosome and vacuole regions and operates in real-time on standard laptops.
      • Riordon et al. (2019) used deep learning to enhance sperm head morphology analysis, achieving a 94.1% true positive rate.
      • Computer-assisted sperm analysis (CASA) with AI (CNNs) allows for precise and quick analysis of sperm trajectories, improving semen analysis accuracy and reducing the need for additional patient information.
      • The Bemaner system (Tsai et al., 2020) uses AI and cloud technology for home-based sperm motility analysis, validated by medical experts.
      • A deep learning model can assess DNA integrity from brightfield images, selecting high-integrity sperm rapidly (under 10 ms).
    • Prediction of Embryo Ploidy Status and Optimal Embryo Selection: Embryo aneuploidy is a significant challenge. AI’s ability to analyze embryologic images and clinical data presents a quicker, more cost-effective, and non-invasive solution compared to traditional PGT-A.
      • A new AI model accurately predicts embryo ploidy from optical microscopy images, with accuracy improving from 65.3% to 77.4% post-optimization. High-scoring embryos by the AI are twice as likely to be euploid.
      • Mixed effects logistic regression achieved an AUC of 0.71 and F1 score of 0.77, confirming blastocyst expansion and trophectoderm quality as key euploidy predictors.
  • Outcome Prediction: Machine learning is refining predictive modeling in IVF, leveraging patient-specific data to enhance the accuracy of clinical outcomes, such as live birth and clinical pregnancy rates.
    • Wang et al. (2022) found that the random forest model (AUC 0.7208) was more effective than logistic regression in predicting clinical pregnancies, identifying ovarian stimulation protocols, female age, and infertility duration as key factors.
    • A hybrid AI model integrating time-lapse video with 31 clinical factors achieved a significant AUC of 0.727 for clinical pregnancy predictions.
    • Tree-based ensemble models like Random Forest and Super Learner excelled in predicting IVF implantation outcomes, with maternal age, embryo transfer timing, gonadotrophin dose, and oestradiol levels being key predictors.
  • AI in Patient Safety and Quality Management: The critical nature of IVF lab procedures demands flawless execution to prevent severe repercussions of human error.
    • ESHRE guidelines advise secondary verification or automated identification for crucial procedures like initial cell identification, embryo transfers, or cryostorage.
    • Electronic witnessing systems (EWS) have effectively prevented sample mix-ups over a decade. Hammer et al. (2022) utilized a CNN that achieved 100% accuracy in identifying patients from their embryo time-lapse images on both day 3 and day 5.
    • A ten-year study on EWS (using RFID and barcodes) tracked over 849,650 points in 109,655 IVF, ICSI, FET, and IUI cycles, achieving a mismatch rate of only 0.251% and critical errors at 0.017%.
  • Micromanipulation Procedures: AI’s advanced image-processing capabilities are revolutionizing embryology by providing precise guidance for intricate procedures such as ICSI and assisted hatching.
    • Jiang et al. (2023) utilized deep learning CNNs to identify critical morphological features in oocytes and embryos, enhancing precision in ICSI and AH procedures with accuracies of 98.9% and 99.41%, respectively.
    • An Automated Intracytoplasmic Sperm Injection (ICSIA) robot demonstrated its effectiveness by automating key IVF steps, achieving a 92.5% survival rate for injected human oocytes and leading to the birth of two infants in a clinical pilot.
  • Social Egg Freezing: This procedure allows women to delay childbearing for non-medical reasons. AI can accurately predict the precise number of oocytes required for successful social egg freezing or fertility preservation by integrating clinical values such as age, AFC, and AMH levels.
    • Kakkar et al. (2023) research suggests that women under 35 should freeze 15 oocytes for a 70% chance of live birth, and 25 for a 95% chance. For women at 34, 10 eggs offer a 75% chance; at 37, 20 eggs are needed; and at 42, 61 eggs are required for the same success rate.
  1. Future Perspectives and Challenges of AI in ART

The integration of AI into ART marks a transformative era, significantly enhancing the accuracy and efficiency of fertility interventions. AI demonstrates extensive utility, from facilitating transvaginal oocyte retrieval (TVOR) by navigating safe pathways, to precisely identifying ideal embryo transfer sites. It particularly revolutionizes micromanipulation tasks in IVF labs by offering levels of precision, consistency, and repeatability that were previously constrained by human variability. The field of AI in embryo transfer, especially in selecting between fresh or frozen embryo transfers, remains relatively unexplored but ripe for research. AI’s ability to forecast odds and success rates based on individual profiles assists in determining the more suitable transfer option. AI also improves communication and trust between patients and clinicians by providing precise, data-driven recommendations.

However, despite these significant technical advancements, practical challenges such as high implementation costs, the need for extensive staff training, and regulatory hurdles are often underexplored. Integrating AI into clinical settings requires substantial investment in both technology and personnel development, as well as navigating complex regulations concerning patient safety and data privacy. Furthermore, AI models depend on large, diverse datasets for training, which can be difficult to obtain given the sensitive and variable nature of reproductive data. Integrating AI into existing clinical workflows also presents challenges, from ensuring system compatibility to maintaining the high levels of accuracy and safety required in patient care.

  1. Conclusion

AI’s role in ART extends to optimizing protocols across various stages of IVF, from ultrasound imaging for follicle assessment to precise gonadotropin dosing and identifying ideal sites for embryo transfers. In the IVF lab, AI excels in micromanipulation techniques, enhancing the precision of procedures such as ICSI and assisted hatching, and it aids clinicians in selecting the optimal embryos for transfer or cryopreservation. The adoption of AI requires a balanced approach, recognizing it as an ally rather than a competitor. This partnership, where AI’s efficiency complements clinicians’ expertise, makes the promise of higher ART success rates attainable. By embracing AI, fertility care professionals can redefine patient outcomes, ensuring the field remains at the forefront of innovation and care excellence.

Introduction to AI in Medicine AI is defined as a branch of engineering that models intelligent behavior to solve complex problems using computers with minimal human intervention. In medicine, AI applications can be broadly categorized into two branches:

  • Virtual Branch: This includes machine learning, which uses mathematical algorithms to improve learning through experience.
  • Physical Branch: This involves robots assisting surgeons during operations or monitoring treatments.

Types of Machine Learning The article identifies four types of machine learning:

  • Supervised Learning: Used when the desired outcome is known, training algorithms on labeled datasets to make future predictions or classifications.
  • Unsupervised Learning: Employed when the target outcome is unknown, allowing algorithms to autonomously process extensive datasets and uncover hidden patterns or structures.
  • Semi-supervised Learning: Particularly useful in medical imaging when both labeled and unlabeled data are present.
  • Reinforcement Learning: The algorithm learns by dynamically interacting with its environment, using a trial-and-error method to achieve specific goals, earning rewards for correct actions and penalties for mistakes.

Common classical machine learning models used in medicine include linear regression, logistic regression, decision trees, random forests (for classification), and Support Vector Machines (SVM) (for regression and classification). Artificial Neural Networks (ANN) represent an evolution of classic machine learning, inspired by biological neural networks, where nodes (neurons) communicate through connections with different weights to achieve a desired outcome. Generative AI, like GPTs, also generates new outputs based on trained data.

Role of AI in Male Infertility Diagnosis AI plays a significant role in improving male infertility diagnostics, offering tools for rapid, accurate, and reproducible analysis:

  • Computer-Aided Sperm Analyzers (CASA): These devices use AI to rapidly analyze semen samples, reducing interoperator variability and providing more accurate measures of sperm motility and kinematics. For example, a t-SNE model can predict fertility with high accuracy using sperm kinetic variables from CASA. However, CASA systems still have limitations in assessing sperm morphology and concentration in complex samples and are not yet a replacement for human operators. The WHO 6th manual (2021) recommends CASA systems as advanced tools primarily for sperm motility and kinematics.
  • Prediction of Male Infertility: Machine learning models have been tested to predict male infertility, with random forest models achieving optimal accuracy (90.47%) and AUC (99.98%) in fertility prediction.
  • Lifestyle and Environmental Factors: AI has been used to evaluate how lifestyle factors like age, alcohol consumption, cigarette smoking, and sedentary lifestyle predict alterations in future fertility, showing them to be more effective predictors than childhood illnesses.
  • Sperm DNA Fragmentation: Algorithms and deep learning methods have been developed to predict sperm with a high DNA fragmentation rate, a known cause of male infertility and Assisted Reproductive Technology (ART) failure.
  • Flow Cytometry: AI tools, such as those integrated into software like FlowJo™ or Cytobank™, analyze flow cytometry data at a single-cell level to provide insights into biofunctional sperm parameters like DNA fragmentation, mitochondrial membrane potential, oxidative stress, and membrane peroxidation.

Use of AI in Assisted Reproductive Technologies (ART) AI is widely used in ART to improve success rates:

  • Embryo Quality Analysis and Selection: Various machine learning and deep learning approaches are employed to analyze embryo quality and assist in selecting the best embryos for transfer.
  • Sperm Selection: Deep learning or SVM methods demonstrate high sensitivity and specificity in selecting spermatozoa with high-quality morphology. AI is also useful in assessing sperm motility with CASA systems. An integrated AI approach considering all these parameters simultaneously could significantly aid embryologists in selecting the best spermatozoa, leading to improved pregnancy and live birth rates.
  • Prediction of Outcomes: An Artificial Neural Network (ANN) trained with 12 features (e.g., woman’s age, gonadotropin dose, endometrial thickness, number of top-quality embryos) showed a cumulative sensitivity of 76.7% and specificity of 73.4% in predicting live births.

Use of AI in Andrological Decision-Making AI assists clinicians in decision-making across several andrological conditions:

  • Varicocele Repair: A random forest model showed high accuracy in predicting patients who would improve their sperm parameters after varicocele repair. Serum FSH levels and bilateral varicocele were identified as fundamental predictive parameters. This approach helps select patients for surgery, potentially reducing overtreatment.
  • Sperm Extraction in Non-Obstructive Azoospermia (NOA): AI algorithms, such as gradient-boosted trees, have been used to identify NOA patients who might benefit from sperm extraction, demonstrating superior accuracy compared to traditional models.
  • Prostate Cancer: AI algorithms based on Convolutional Neural Networks (CNN) have improved the standardization of the Gleason score in prostate biopsies, showing high precision in detecting tumor areas and assigning Gleason patterns with accuracy similar to pathologists, thus reducing interoperator variability.
  • Erectile Dysfunction (ED):
    • A Clinical Decision Support System (CDSS) based on an integrated genetic algorithm and SVM demonstrated high sensitivity, specificity, and accuracy in predicting the incidence of ED, using features like age, comorbidities, and related variables.
    • AI has contributed to developing new questionnaires and evaluating ED through medical imaging. For instance, a visual scale questionnaire assessed ED severity with superior accuracy compared to more complex ones like IIEF-5.

Use of AI in Andrological Imaging AI applications in andrological diagnostic imaging are noteworthy:

  • Prostate Cancer: Machine learning models using MRI data (perfusion maps, apparent diffusion coefficient, T2 signal intensities) can predict the presence of clinically significant prostate cancer (Gleason score ≥3+4) more effectively than traditional systems. AI is useful in segmentation, lesion detection, and aggressiveness prediction.
  • Testicular Function: AI has confirmed that testicular echo structure reflects reproductive function. A radiomics approach to testicular echotexture can predict testicular spermatogenic capacity and correlate with pituitary function, highlighting the importance of ultrasound diagnostics in male fertility.
  • Erectile Dysfunction: An SVM model demonstrated that diffusion tensor imaging indices of certain brain areas, evaluated by MRI, could aid in diagnosing ED caused by altered veno-occlusive mechanisms.

Use of AI in Andrological Surgery The physical branch of AI, specifically robotics, plays a role in andrological surgery:

  • Robotic-Assisted Surgery: Augmented reality techniques for prostate reconstruction, based on multiparametric MRI images, are being tested to improve outcomes in prostatectomy by tailoring procedures to the patient.
  • Microsurgical Procedures: Robotics could be useful for vasectomy reversal, varicocelectomy, testicular sperm extraction, and spermatic cord denervation in infertility treatment.
  • Advantages: Robotics offers advantages such as the elimination of operator tremors, three-dimensional visualization, and a reduced need for qualified surgical assistance.
  • Limitations: Despite the benefits, limitations include a lack of sufficient studies demonstrating superiority, skepticism from experienced surgeons regarding delicate tissue manipulation, and the high costs compared to traditional surgery.

Strengths and Limitations of AI in Medicine The article outlines several strengths and limitations of AI in medicine: Strengths:

  • Increased Precision and Accuracy: AI can provide more precise and accurate diagnoses, resolving cases where physicians might disagree.
  • Enhanced Diagnostic Performance: It can increase the diagnostic sensitivity and specificity of tests.
  • Assistance in Decision-Making: AI aids physicians in disease management, potentially reducing medical errors, costs, morbidity, and mortality.
  • Improved Healthcare Access: Deployment in rural areas could help overcome the lack of expert physicians, bridging the gap between rural and urban healthcare quality.

Limitations:

  • High Costs: The hardware required to manage AI algorithms and robots used in surgery can be very expensive.
  • Safety and Reproducibility Concerns: Questions remain regarding the safety and reproducibility of AI software.
  • Ethical and Legal Issues: Protecting patient privacy poses significant ethical and legal challenges.
  • Deskilling of Clinicians: Over-reliance on AI could potentially reduce clinicians’ decision-making skills.
  • Limited Generalizability: Many algorithms have been tested on limited populations, raising questions about their effectiveness when applied to the general population.

Future Perspectives on AI in Andrology AI is expected to enable increasingly precise execution of ART procedures in the future:

  • Personalized ART Protocols: This includes optimizing stimulation protocols, more accurate selection of gametes, and identifying the best embryo for transfer, all aimed at improving pregnancy and live birth rates.
  • NOA Patient Identification: AI could help identify NOA patients who would benefit from sperm retrieval procedures using biomarkers like leptin and FSH.
  • Genetic Testing for Azoospermia: AI could assist in identifying azoospermic patients who require further genetic testing.
  • Objective Predictive Algorithms: Developing more objective and specific predictive algorithms will become a useful tool for clinicians, assisting in clinical decision-making, reducing care costs, and enabling patient-tailored medicine.

Conclusion The evidence suggests a growing role for AI in supporting physicians in managing andrological problems, particularly in improving ART success rates, aiding physician decision-making, and enhancing diagnostic performance through integration with imaging. While promising, many studies are still needed to validate AI’s effectiveness, especially through properly designed multicenter studies on large populations to confirm reliability and validity. The future task for physicians will be to work in an integrated manner with AI, rather than passively relying on it, to ensure continued professional competence and maintain the fundamental prerequisites of trust and empathy in patient care.

Introduction to AI in ART: Promise vs. Reality

Infertility is a noteworthy concern in the contemporary era, with a global estimated prevalence of one in six couples struggling with it. Despite significant technological advancements in IVF, such as extended culture to the blastocyst stage and embryo vitrification which have improved implantation and cryopreservation outcomes, key performance indicators like fertilization and embryo survival rates have seen limited improvement. This stagnation is attributed to fragmented practices and insufficient standardization across IVF laboratories, underscoring the need for integration with “dry sciences” like AI.

AI holds immense potential to optimize clinical workflows and address inefficiencies in IVF. However, the adoption of advanced technologies like electronic witnessing systems and robust cryostorage tracking remains low, with the field still heavily reliant on paper records and manual processes. While the global AI market is projected for significant growth, its penetration in IVF remains modest compared to other medical sectors. Despite concerns regarding ethics, efficiency, and acceptance, most embryologists view the integration of AI as inevitable and potentially beneficial.

How AI in Medicine Inspires Applications in Reproductive Medicine

AI’s evolution from decision-making models to its integration into healthcare has revolutionized clinical reasoning, offering objectivity over human biases and experience. AI currently impacts healthcare through rapid image interpretation, workflow efficiency, and patient empowerment via applications and wearables. These innovations are increasingly relevant to ART, where AI-powered tools can educate patients via chatbots and manage secure documentation.

Embryologists face increasing workplace complexity, staff shortages, and data management challenges, making AI an indispensable tool for improving workflow, standardization, and decision-making. Just as automated systems track driver fatigue, AI can assist embryologists by reducing distractions and enhancing precision. AI’s success in radiology and pathology, particularly in image interpretation and standardization, highlights its potential for addressing similar challenges in IVF, especially in morphological assessments. AI also extends to precision tasks in robotic surgery and single-cell procedures on gametes and embryos, relying on automation to enhance outcomes and reduce variability. Beyond image analysis, AI can optimize dose regimens and predict adverse drug events, with similar approaches now being applied in ART to personalize ovarian stimulation protocols and improve cycle management.

AI research in reproductive medicine has seen a modest but steady incline in publications over the past few years. However, a significant proportion of these publications are reviews or commentaries rather than original research, and many studies lack a clear emphasis on outcome-driven differences or tangible clinical applications, limiting their relevance to real-world IVF practices.

AI Applications in ART: Key Domains

The article delineates AI’s potential applications across various stages of the IVF procedure:

  • AI in Clinic and Cycle Management:
    • Patient Selection and Management: AI systems like Fenomatch use facial biometrics to help patients select donors.
    • Clinical Result Management: PGTai uses AI to provide precise chromosomal profiles from blastocyst biopsies, minimizing noise in next-generation sequencing data and improving the reliability of embryo selection.
    • Workflow Optimization: Companies like Cercle.AI provide AI-driven data pooling solutions for image processing, patient onboarding, treatment planning, and predictive analytics. FertilAI offers automated cycle management tools to optimize trigger timing and frozen embryo transfer scheduling, streamlining workloads.
    • Patient Preparation: Future Fertility has developed tools to predict implantation by evaluating endometrial thickness and receptivity with greater accuracy. Folliscan uses deep learning to monitor follicular development via ultrasound, providing recommendations for trigger timing and reducing procedure time.
  • AI in Andrology (Semen Analysis & Sperm Selection):
    • AI has revolutionized semen analysis, building upon Computer-Assisted Sperm Assessment (CASA). It enhances sperm selection by reducing human error and offering objective and standardized assessments of sperm quality.
    • AI-powered tools excel at classifying sperm morphology, identifying subtle defects with high accuracy, with some automated classifiers achieving up to 94% accuracy rates.
    • There is also promise for future DNA fragmentation analysis to identify viable sperm with intact genetic material.
    • Narrow AI is already commercially used in CASA systems to evaluate sperm kinematics, debris, and agglutination.
    • For Intracytoplasmic Sperm Injection (ICSI), AI tools like SiD (Sperm Identification) offer real-time sperm selection, correlating motility parameters with improved fertilization and blastocyst rates, showing comparable results to experienced embryologists.
    • AI is also advancing sperm identification in testicular samples, reducing search times for embryologists during micro-testicular sperm extraction procedures.
  • AI in Oocyte Selection:
    • AI marks a shift towards objective, standardized methods for oocyte assessment, enhancing precision and reproducibility in IVF.
    • AI-driven algorithms are pivotal in analyzing digital oocyte images, offering promising predictions of oocyte viability. An example is Violet, an AI algorithm with up to 90% accuracy in predicting fertilization outcomes from a single image, by analyzing subtle features like dimensions, shape, anomalies, and zona pellucida characteristics.
    • AI also aids specialized procedures like ICSI by identifying optimal injection sites, guiding tool positioning, and confirming membrane breakage, thus reducing damage risk during needle punctures.
  • Innovating Embryo Selection:
    • AI-based tools offer a standardized, non-invasive approach to prioritizing high-quality embryos for transfer. Models trained on large datasets analyze morphological features to estimate the likelihood of pregnancy.
    • ERICA, an AI system trained on day 5 embryo images, predicts ploidy potential and probability of pregnancy, with studies suggesting it may surpass human embryologists in accuracy.
    • Time-lapse incubators support AI-driven embryo selection by providing continuous, non-disruptive monitoring.
    • AI tools like KIDScore, Eeva, Life Whisperer, and FertilAI provide decision support for embryo selection. Retrospective studies link time-lapse parameters with embryo viability, aneuploidy status, and pregnancy rates.
    • However, despite AI reducing assessment time and standardizing evaluations, current evidence does not conclusively show better clinical outcomes compared with traditional methods. Some critics warn against overhyping these technologies. A randomized controlled trial found AI to be non-inferior to manual methods but noted its speed advantage and reduced variability. Large prospective trials are still needed, but AI-assisted embryo selection could potentially reduce the number of embryos transferred and improve first-transfer success rates.
  • AI and Automation of ART:
    • Automation aims to address the shortage of embryology expertise by reducing reliance on traditional tasks, allowing embryologists to focus on higher-level responsibilities.
    • Two main approaches to automation include microfluidics for autonomous IVF processes (e.g., sperm preparation, vitrification) and modeling automation on embryologists’ activities (e.g., pipetting, fluid handling).
    • AI, with computer vision, deep learning, and neural networks, enhances automation by improving efficiency, accuracy, and consistency, and reducing variability between IVF centers and operators. This standardization may make fertility care more affordable and accessible.
    • Automated ICSI robots have successfully injected human oocytes, leading to births from partially automated ICSI. Other AI-powered robotic systems have performed remote automated ICSI with comparable outcomes to manual methods.
    • Automation is expanding into sperm preparation, oocyte retrieval and denudation, dish preparation, and cryopreservation, with several startups developing and commercializing these technologies.

AI in Research

IVF laboratories generate vast amounts of data, though it’s often difficult to access or share. AI systems are expected to produce increasingly accurate classifications by accessing “big data,” potentially offering answers for patients without clear diagnoses and recommending specific therapies or optimal timeframes for critical tasks like ovarian stimulation or embryo transfer. AI is increasingly used in IVF research to correlate -omics data with clinical outcomes, enabling a deeper understanding of genetic and biochemical factors influencing embryo viability, ovarian response, and implantation. However, the full promise of this field remains unrealized. It’s crucial for AI systems to provide explainability, offering insights into what the model is analyzing, enabling AI-empowered decisions. The deployment of automation technologies could standardize processes and outcomes across sites, facilitating the collection of consistent datasets.

Drawbacks and Limitations of AI in Medicine and IVF

Despite its potential, AI adoption in IVF remains slow due to several drawbacks:

  • High Implementation Costs: Significant investment in technology and personnel development is required.
  • Safety and Reproducibility Concerns: Questions persist regarding the safety and reliability of AI software.
  • Data Security and Privacy: AI’s dependence on large datasets makes it vulnerable to breaches, raising notable risks, especially in regions with inconsistent privacy laws. Protecting patient privacy poses significant ethical and legal challenges.
  • Bias in AI Models: AI systems may perpetuate disparities if trained on non-representative datasets, potentially impacting outcomes for specific populations.
  • Erosion of Expertise: Over-reliance on AI could potentially reduce clinicians’ and embryologists’ decision-making skills. A co-convergence of human and AI inputs is essential for quality management.
  • Transparency (“Black Box” Problem): Many AI systems operate as “black boxes,” which could erode trust and lead to errors if outputs are misinterpreted or blindly adopted.
  • Ethical Concerns: AI’s role in embryo selection and genetic predictions raises questions about “designer babies” and may lead to societal pushback, complicating patient counseling and informed consent.
  • Integration Challenges: Diverse IVF protocols and equipment can hinder AI integration, causing inefficiencies or errors.
  • Limited Generalizability: Many algorithms have been tested on limited populations, making their effectiveness when applied to the general population questionable, thus reducing validity and reliability.
  • Sustainability Concerns: The rapid expansion of data centers required for AI’s massive databases and energy-intensive computational processes raises significant environmental concerns, with estimated electrical consumption increasing dramatically.

Regulation in AI

AI in healthcare presents significant regulatory challenges, balancing safety with innovation. The WHO highlights interconnected challenges, including patient consent, data protection, and cybersecurity. While frameworks like HIPAA and GDPR address data privacy, gaps exist for unregulated data, and cyberattacks increasingly threaten healthcare institutions. Legal liability remains a concern due to AI’s opacity, complicating the attribution of responsibility.

Globally, AI regulation for Software as a Medical Device (SaMD) struggles to adapt to autonomous, evolving technologies. Progress in harmonizing standards is slow, despite efforts like the US-EU Trade and Technology Council’s voluntary AI code of conduct. Leading jurisdictions have adopted provisions for good machine learning practices, but inconsistent terminology and definitions hinder global integration. The FDA has approved nearly 1000 AI-enabled devices, primarily in radiology and cardiology, but adaptive AI systems challenge traditional regulatory pathways. The EU AI Act, the first major comprehensive AI regulation, emphasizes patient safety, data governance, and ethical use. The UK’s Medicines and Healthcare products Regulatory Agency has improved post-market surveillance.

Despite advancements, regulatory gaps persist, as frameworks often lag behind AI’s rapid evolution, and global harmonization remains limited. Ethical concerns require adaptive and collaborative solutions. The Croatia Consensus offers best practices for validating AI in ART, emphasizing standardized data collection, robust privacy measures, staff training, and ethical use.

Conclusion and Future Outlook

The widespread implementation of AI in IVF remains elusive, despite the clear and pressing need for tools to enhance the effectiveness, consistency, and accuracy of clinicians’ and embryologists’ work. This slow adoption reflects that, without demonstrable and significant advantages, new technology in IVF is rarely adopted rapidly. Historically, it takes at least 7-10 years for new medical technologies to reach financial equilibrium and often a generation for true acceptance.

For AI to truly transform IVF, it must not only perform but also inspire trust, integrate seamlessly into daily workflows, and deliver measurable improvements in patient outcomes, achieving more than just non-inferiority. Simplicity, accuracy, and affordability are non-negotiable. While AI solutions hold considerable scientific promise and persistent hype, they have yet to consistently demonstrate the tangible benefits needed to justify systematic adoption. The path forward demands transparency, clear regulations, and an intuitive fit within current IVF laboratory practices.

Real-world benefits, such as faster, more accurate results, improved outcomes, and increased IVF success, must be evident. The human element—clinicians and embryologists—remains indispensable. Without their willingness to embrace AI, share high-quality data, and collaborate in refining these tools, AI’s progress in ART will stall. Trust, collaboration, and a shared vision for the future are vital for AI to achieve its full potential as a transformative force in reproductive medicine. The future hinges on balancing technological innovation with the needs, expectations, and expertise of the professionals it aims to support.

Methodology

The study utilized a retrospective analysis of data collected between 2019 and 2022 from 11,988 embryos derived from 2,666 egg retrievals across 11 fertility centers in France, Spain, and Morocco.

  • Data Sources: The data included videos of embryos recorded using various time-lapse systems (TLS), specifically Embryoscope, Embryoscope+, GERI, or MIRI. Embryologists provided information on whether embryos were discarded, frozen, or transferred (fresh or frozen), along with transfer dates and outcomes (clinical pregnancy/fetal heartbeat, no heartbeat, or live birth). Clinical variables were collected at the patient level at the time of egg retrieval.
  • Ethical Considerations: The study used retrospective and de-identified data, meaning it was exempt from ethical review and the requirement for informed consent.
  • EMBRYOLY Algorithm:
    • The EMBRYOLY algorithm is a modified version of a previous model, employing a UniFormer, which is a transformer-based architecture specifically designed for video processing.
    • It contains 21 million parameters and was pre-trained on the Kinetics open-source video dataset, which covers 400 human action classes.
    • The model’s inputs are sequences of time-lapse images from the central focal plane of the TLS.
    • Its output is a numerical score between 0 and 1, indicating the likelihood of a fetal heartbeat, which correlates with the likelihood of clinical pregnancy.
    • Crucially, the algorithm was trained without any human-annotated data and only the computer vision component was used, excluding clinical features, to focus purely on embryo ranking.
  • Experimental Design: Three distinct experiments were conducted on different subsets of the data:
    1. Correlation Analysis: This evaluated the statistical association between EMBRYOLY’s score and known transfer outcomes (clinical pregnancy and live birth) for all transferred embryos (2,657 for clinical pregnancy, 2,121 for live birth), encompassing both cleavage and blastocyst stages. This also included hold-out analyses using data from two independent clinics (one French, one Moroccan) whose data had not been used to train the algorithm.
    2. Agreement Rate: This assessed how often EMBRYOLY, when used as a standalone tool, would have ranked the same embryo as the embryologist’s first choice in cohorts where multiple embryos were available for transfer but only one was actually transferred (926 egg retrievals, 4,974 embryos).
    3. Impact on Cycles to Pregnancy (CTP) and First Cycle Pregnancy Rate (FCP): This was the main analysis for EMBRYOLY as an adjunct tool. It focused on cohorts with multiple single blastocyst transfers where at least one transfer resulted in a clinical pregnancy and one did not (111 egg retrievals, 260 embryos).
      • Key Hypotheses: It was assumed that embryologists initially transferred the embryo they considered most viable. Additionally, it was hypothesized that the transfer outcome would remain consistent regardless of whether the embryo was transferred fresh or frozen, or the order of transfer for a given patient.
      • For this analysis, EMBRYOLY’s ranking was applied only to the embryos that embryologists had initially chosen to transfer, reflecting its use as a supportive tool rather than a full replacement.
  • Statistical Methods: Logistic regressions were used for correlation analysis. McNemar’s test was applied to compare FCPs, and Wilcoxon tests were used for CTP comparisons, with a statistical significance threshold of p < 0.05.

Key Results

  • Patient Demographics: The average patient age was 34.1 ± 4.6 years, with an average BMI of 24.1 ± 4.5.
  • EMBRYOLY Score Correlation:
    • For blastocyst transfers, a 0.01 increase in EMBRYOLY’s score was statistically associated with a 2.8% relative increase in the likelihood of clinical pregnancy (Odds Ratio (OR) 1.028, 95% Confidence Interval (CI) 1.023-1.033, p < 0.001).
    • A significant correlation was also found for live births from blastocyst transfers (OR 1.024, 95% CI 1.016-1.032, p < 0.001).
    • For Day 3 cleavage embryos, the score significantly correlated with clinical pregnancy (OR 1.026, 95% CI 1.017-1.036, p < 0.001) and live birth (OR 1.019, 95% CI 1.004-1.035, p = 0.01).
  • Generalizability (Hold-out Clinics): The positive correlation for clinical pregnancy was maintained in 7 out of 9 eligible clinics, including the two held-out clinics (French clinic: OR 1.022, p = 0.02; Moroccan clinic: OR 1.026, p = 0.03). The live birth correlation also held true for 5 out of 8 clinics, including the Moroccan hold-out.
  • Agreement Rate (Standalone Use): When EMBRYOLY acted as a standalone ranking tool, its highest-ranked embryo matched the embryologist’s chosen embryo in 62.2% of egg retrievals (576 out of 926), which was significantly higher than a random agreement rate of 18.6%.
  • Impact of Adjunct Use on Outcomes:
    • The hypothetical first cycle clinical pregnancy rate (FCP) could have increased from 19.8% (embryologists alone) to 44.1% (EMBRYOLY + embryologists) (McNemar’s test: P < 0.001).
    • This improvement could have reduced the average number of cycles to clinical pregnancy (CTP) from 2.01 to 1.66 (Wilcoxon test: P < 0.001).

Discussion

The study highlights its novelty as the first to evaluate an AI-powered embryo evaluation software as an adjunct tool for embryologists, focusing on its potential to reduce the number of cycles to pregnancy when integrated with existing expertise. Previous AI studies in this field often trained algorithms for binary outcomes (e.g., pregnancy vs. no pregnancy) and did not conclusively demonstrate their ability to rank within restricted sibling embryo cohorts.

  • Comparison to Prior Research:
    • The study differentiates itself from work like Diakiw et al. (2022), which evaluated AI on “fictional cohorts” and assumed blind adherence to AI recommendations, without directly comparing to embryologist ranking.
    • It also contrasts with Cimadomo et al. (2023), who assessed iDAScore v1.0 as a standalone tool on PGT-A biopsied embryos from a single center. This prior study’s approach limited generalizability and made it difficult to quantify the full impact on live birth rates if the AI selected a non-transferred embryo. The current study’s “adjunct use” model, by focusing only on embryos considered for transfer by embryologists, circumvents this issue and allows for more accurate quantification of impact.
    • Furthermore, the current study’s dataset included over 98% non-biopsied embryos, making its findings highly relevant for IVF centers that do not routinely perform PGT-A, unlike Cimadomo et al.’s focus solely on euploid embryos.
  • Advantages of EMBRYOLY:
    • The results indicate that EMBRYOLY can effectively rank embryos based on their potential for clinical pregnancy and live birth, even for cleavage-stage embryos. This ability to rank Day 3 embryos is particularly useful for clinics that perform early transfers.
    • The study demonstrated EMBRYOLY’s generalizability, with statistically significant results in new, unseen clinics, including those with different patient demographics (e.g., younger average age and higher BMI in the Moroccan clinic).
    • EMBRYOLY offers a pathway to reduce subjectivity and standardize practices across different centers and countries. Unlike human embryologists whose criteria can vary, an algorithm provides reproducible recommendations based on objective clinical outcomes.
  • Limitations of the Study:
    • The study relied on the hypothesis that embryologists always transfer the embryo they deem most viable first, and that the transfer outcome is independent of the fresh/frozen status or transfer order (assuming consistent uterine and endometrial environment).
    • It also assumed that embryologists would strictly follow EMBRYOLY’s recommendations after their initial pre-selection, which may not always be the case in real-world practice.
    • There’s a potential for data selection bias if clinics primarily uploaded data from good-quality embryos, although the multi-centric nature of the study helps to mitigate this.
    • The comparison was limited to the decisions of individual embryologists, not accounting for potential variability among different embryologists, although a diverse group of over 11 embryologists with varying experience levels was involved.
  • Overall Impact: The study suggests that EMBRYOLY can significantly increase the first cycle clinical pregnancy rate and minimize the time to clinical pregnancy for patients. This could potentially lead to higher cumulative success rates by reducing patient dropout after initial failed treatments. It may also encourage fresh transfers for suitable patient populations and lessen the physical and emotional burden of repeated failed cycles. For IVF centers, such a tool could make embryo evaluation more streamlined and consistent.

Conclusion

In conclusion, this study strongly suggests that EMBRYOLY has the potential to help embryologists effectively identify the most promising embryos within a cohort across various time-lapse systems, IVF centers, and countries. By providing an improved ranking, the tool aims to increase the first cycle clinical pregnancy rate and reduce the time patients need to achieve clinical pregnancy, thereby enhancing the efficiency and emotional well-being associated with fertility treatments.

Background and Context

ART has seen significant evolution, leading to the birth of over 10 million babies, yet the live birth rate remains around 30-40%, indicating a persistent need for improvement. Key factors influencing successful pregnancies include the in vitro culture of high-quality embryos, the accurate selection of embryos with the highest implantation potential, and synchronization with the optimal uterine implantation window. Traditional morphological assessment of cleavage-stage embryos often has limited predictive value for their developmental potential, leading to implantation rates of only 20-40%. This subjectivity and inconsistency can make it challenging for embryologists to select the best embryo, especially when many high-quality options are available.

The debate continues regarding the effectiveness of time-lapse culture alone in improving pregnancy or live birth rates. However, this study posits that the combination of time-lapse culture with AI could improve pregnancy rates in fresh cycles. The study also highlights the advantages of single cleavage-stage transfers, which can effectively reduce multiple birth rates and associated risks like miscarriages and perinatal complications (e.g., preterm births, low-birth-weight babies), issues often stemming from the common practice in developing countries of transferring multiple embryos to achieve higher pregnancy rates. While single embryo transfers are internationally recommended, primarily for blastocysts due to their high potential, fresh blastocyst transfers carry risks such as increased cancellation rates, reduced pregnancy rates due to early endometrial implantation window closure with certain ovulation regimens (e.g., antagonists), potential long-term effects on offspring (e.g., shortened telomeres), and sex ratio imbalances. Single cleavage-stage transfers are presented as a means to mitigate these concerns.

Methodology

The study conducted a retrospective analysis of 105 fresh embryo transfer cycles performed between August 2023 and March 2024 at the Reproductive Center of the Affiliated Hospital of Guangdong Medical University.

  • Participants: Patients included were 38 years old or younger, underwent long agonist or antagonist protocols, had a fresh transfer on day 3 of a single cleavage-stage embryo cycle, and had embryos cultured in an EmbryoScope+ time-lapse system. Exclusion criteria included reproductive system abnormalities, chromosomal anomalies, a history of uterine surgery, or missing data. No statistically significant differences were found in baseline characteristics (age, AMH, AFC, basal sex hormones) or ovulation program distribution among the study groups.
  • Embryo Culture and AI Scoring: All embryos were cultured using time-lapse technology. The AI model used for scoring was iDAScore V2.0.
    • The iDAScore AI model is a deep learning and neural network-based system trained to analyze sequences of time-lapse images.
    • Its inputs were solely images from time-lapse sequences, and its outputs were numerical scores ranging from 1.0 to 8.0 (for Day 3 Models) which correlated with the likelihood of a fetal heartbeat (FHB).
    • Crucially, the iDAScore did not use any human-annotated data for its training.
  • Embryo Categorization: Embryos were categorized into three groups based on their iDAScore V2.0 and cell count:
    • Group A: 8 cells, iDA: 1.0-5.7.
    • Group B: 8 cells, iDA: 5.8-8.0.
    • Group C: >8 cells, iDA: 5.8-8.0.
  • Clinical Procedures:
    • Controlled Ovarian Stimulation (COS): Patients received either long GnRH agonist or GnRH antagonist protocols. HCG was administered based on follicle diameter and E2 levels.
    • Fertilization and Culture: IVF or ICSI was performed. Oocytes showing a second polar body were cultured individually in EmbryoScope+ time-lapse incubators. Embryo grading referred to the Istanbul Consensus.
    • Embryo Transfer: Selection was based primarily on the iDAScore 2.0. A serum hCG level > 5 U/L indicated a positive pregnancy, and an intrauterine pregnancy with early cardiac motion confirmed a clinical pregnancy.
  • Statistical Analysis: SPSS 26.0 was used, employing c2 tests for enumeration data, independent sample t-tests for normally distributed data, and LSD t-tests for pairwise multiple comparisons. A p-value < 0.05 was considered statistically significant.

Key Results

  • iDAScore Distribution: The iDAScores were significantly higher in Group C (7.3 ± 0.5) compared to Group B (6.7 ± 0.5), and also significantly higher in Group B (6.7 ± 0.5) compared to Group A (4.8 ± 1.0) (p < 0.001).
  • Embryo Quality and Development:
    • The mean number of high-quality embryos was highest in Group C (4.7 ± 3.0), followed by Group B (3.6 ± 1.7), and Group A (2.1 ± 1.2) (p < 0.001).
    • The number of blastocysts formed was higher in Group B (4.1 ± 2.2) than in Group A (2.5 ± 2.4) (p = 0.009).
    • The rate of blastocyst formation was also higher in Group B (56.4 ± 28.4%) compared to Group A (37.9 ± 25.0%) (p = 0.008).
  • Ongoing Pregnancy Rates: There was no statistically significant difference (p = 0.392) in the ongoing pregnancy rate for single cleavage-stage transfers between Group B (54.5%, 30/55) and Group A (38.1%, 8/21), though there was a tendency for Group B to be higher. Group C also showed a higher clinical pregnancy rate (69.0%) and ongoing pregnancy rate (55.2%) compared to Group A and B, but these differences were not statistically significant when compared directly.
  • Predictive Value of iDAScores: iDAScores proved to be reliable predictors of the probability of embryo development into blastocysts. For instance, embryos with iDAscores above 5.7 (e.g., Well 10: iDA=6.8, Well 5: iDA=5.7, Well 13: iDA=5.7) developed into good-quality blastocysts, whereas those below 5.7 (e.g., Well 2, Well 6, Well 9) did not form usable blastocysts by Day 5.

Discussion and Conclusion

The study concludes that combining time-lapse culture with AI scoring may enhance ongoing pregnancy rates in single cleavage-stage fresh transfer cycles. This approach aims to address the challenge faced by embryologists in ranking cleavage-stage embryos, where traditional morphological assessment can be subjective and inconsistent. By providing an objective scoring system based on developmental dynamics, AI can potentially outperform traditional methods.

The ability of iDAScores to predict the developmental potential of 8-cell cleavage stage embryos, even when they appear morphologically similar, is a significant finding. This suggests that embryos with higher iDAScores have a higher potential for developing into blastocysts and for ongoing pregnancy. This capability allows embryologists to make more informed decisions regarding transfer, freezing, or continuing to culture embryos based on the patient’s overall embryo score. For patients with high iDAScores, a single cleavage-stage embryo transfer could be chosen, while for those with lower scores, the number of embryos transferred could be adjusted based on individual patient factors. This personalized approach could simultaneously maintain pregnancy rates and reduce multiple births, benefiting the reputation of reproductive centers.

Furthermore, the integration of time-lapse culture with AI scoring significantly reduces the workload of embryologists by automating the evaluation of kinetic parameters from large amounts of embryo image data. This efficiency helps embryologists select embryos with the highest developmental potential, streamlining the decision-making process for transfer sequencing, especially when multiple embryos are available.

The study acknowledges practical considerations, such as the high cost of time-lapse incubators and specialized petri dishes, which may limit widespread adoption in developing countries. However, it suggests prioritizing the use of available time-lapse incubators for patients undergoing fresh cycle transfers to optimize pregnancy rates. For other patients, transferring blastocysts in a freeze-thaw cycle can still maintain high pregnancy rates, as extended culture effectively screens out embryos with low developmental potential.

In essence, the study advocates for the use of time-lapse culture combined with AI as a non-invasive, rapid, and objective method to guide embryo selection, particularly for single cleavage-stage fresh transfers. It aims to shorten the time to first pregnancy, ensuring pregnancy rates while minimizing the risks associated with blastocyst transfer and the challenges of subjective human assessment.

The study highlights that despite considerable technological advancements in assisted reproduction, current methods for selecting oocytes to fertilize or freeze still heavily rely on subjective human interpretation and evaluation, and there is no standardized scoring method for oocyte quality, especially regarding its live birth potential. The authors propose that AI and machine learning could significantly improve outcomes by learning from data and assisting embryologists in their decision-making process, leading to more consistent results.

Methodology of the Review

The systematic review followed established guidelines, including the PRISMA guidelines.

  • Research Questions: Two primary research questions guided the analysis:
    • R1: What features of oocytes, considered by embryologists in medical practice to classify them or assess their capabilities in enhancing IVF outcomes, can be leveraged by AI applications? This question aims to identify oocyte attributes fundamental for automating evaluation and understanding valuable factors for classification, potentially revealing research gaps for future AI models.
    • R2: What are the applications, potential, and limitations of AI-based techniques when applied to oocyte assessment in IVF procedures? This question explores how AI methods can improve efficiency and accuracy, reduce biases, and addresses ethical concerns related to AI-backed decision-making in medical practice.
  • Queries and Sources: Formalized queries were developed and used to search two scientific databases: Scopus and PubMed.
    • Initial search retrieved 623 articles from Scopus and 80 from PubMed.
  • Screening Process:
    • Articles not in English and 28 duplicates were excluded, reducing the total to 648.
    • Two successive rounds of screening on titles and abstracts were performed to ensure relevance to the two research questions.
    • This process yielded 57 relevant articles (47 from Scopus, 10 from PubMed) for the review.
  • Quality Assessment: Three quality assessment questions (A1, A2, A3) were formulated to evaluate the scholarly value of each article, with scores from 0 to 2 for each question:
    • A1: Does the research describe and provide enough information to reproduce the experiments?
    • A2: Does the study’s analysis and interpretation of its findings, supported by a dataset of sufficient size, effectively address the research questions and objectives?
    • A3: How much does the use of AI-based technology contribute to the overall findings and conclusions, and how central is it to the research methodology and analysis?
    • Articles with a total quality assessment score lower than 4 were discarded.
  • Results of Screening: Out of 57 articles, only 12 original research articles met the required quality threshold and served as the primary source for the systematic literature analysis. The earliest relevant article dated back to 2010, with a significant upward growth in interest over the last four years.

Data Sources and Associated Challenges

AI-based techniques in IVF primarily rely on data sources such as still images, time-lapses, and clinical data.

  • Lack of Standardization: Image capture is often not standardized across IVF clinics, which use different microscopes, lenses, and cameras, leading to potential biases in datasets.
  • Low Sampling Rate: Time-lapse incubators, designed for slow embryo dynamics, typically capture frames every 8 minutes. This low sampling rate can render video-based algorithms ineffective for studying oocytes, which have faster dynamics.
  • Data Biases: Datasets often originate from single clinics, meaning patients might receive the same infertility treatment, which can introduce difficult-to-model biases.
  • Dataset Size: A major challenge is the availability of high-quality, annotated datasets for training AI models, as most reviewed articles used small datasets (e.g., 34 images, 93 videos).
    • Many datasets are unbalanced (more samples from one class than others), leading to models with good accuracy on over-represented classes but poor accuracy on others.
    • Small datasets limit the models’ ability to capture feature variability and generalize.
  • Reproducibility: Most articles provide only high-level descriptions of machine learning models, making it challenging to reproduce results and understand methodologies, with only one reviewed article making its code publicly available.

Oocyte Characteristics Leveraged by AI (Research Question R1)

The review identified predominant groups of oocyte characteristics that embryologists consider for quality assessment and which AI applications can utilize:

  • Clinical Data: Information gathered from patients, such as female age, which is a strong indicator of oocyte live birth potential. AI can help uncover hidden patterns and relationships within these factors.
  • Morphometric Parameters: These are morphological features of egg cells or their structure. They are commonly used to assess the quality and live birth potential of oocytes and embryos. Examples include:
    • Oocyte diameter: Believed to relate to a woman’s age.
    • Zona pellucida: Its texture, thickness, and size are considered descriptive of oocyte quality.
    • Cytoplasmic granularity.
    • Spindle position and visibility.
    • The relevance of these features in predicting live birth potential is still debated, but they are considered correlates of oocyte quality.
  • Morphokinetic Parameters: Time-based measurements and characteristics of embryonic cell development observed during early stages, typically derived from time-lapse imaging. These include:
    • Time of nuclear maturation, polar body extrusion, metaphase II or second polar body extrusion.
    • Pronuclear formation, cleavage timing, and blastocyst formation and timing.
    • While time-lapse monitoring aids in embryo selection, reducing manual evaluation, this occurs at the embryo level, often after the oocyte has been fertilized.
  • Cytoplasmic Movement: Intrinsic dynamics of an egg cell, distinct from morphokinetic parameters.
    • Sperm entry triggers rapid cyclic cytoplasmic movements associated with calcium oscillations, observable with fast sampling rates (around 10 seconds).
    • However, much of the reviewed literature studies cytoplasmic movements at the oocyte level using lower sampling rates (in minutes), which may be insufficient to capture relevant dynamics, leading to a lack of comprehensive understanding.

AI-Based Approaches for IVF (Research Question R2)

AI-based techniques have been employed in four significant categories to address challenges in ART:

  • Segmentation: This is a fundamental image processing step to partition an image into regions representing distinct objects (e.g., oocytes and their components).
    • Early methods used classical computer vision (e.g., contrast adjustment, noise reduction, contour detection).
    • More recent approaches utilize deep learning, such as convolutional neural networks (CNNs), to autonomously detect cytoplasm circumference.
    • Challenges include varying camera/brightness settings across clinics and the need for large, robust datasets.
  • Feature Extraction: This process reduces raw data (images, time-lapses, clinical data) into relevant and informative features for machine learning algorithms.
    • Examples include morphological aspects (dimensions, circularity, symmetry), textural aspects (local binary patterns), and time-related aspects (developmental stages, age, hormonal data).
    • Novel approaches include correlating oocyte mechanical properties with live birth potential using physically informed digital models.
    • Cytoplasmic movement data is extracted using Particle Image Velocimetry (PIV), though the reviewed literature often uses PIV without proper consideration of all necessary components (e.g., adequate seeding, illumination) and often with insufficient sampling rates.
  • Classification: This involves evaluating relationships among objects based on their features to categorize them into predefined classes (e.g., oocytes with high live birth potential).
    • The most studied phase, often using binary classifiers (Support Vector Machines, Neural Networks, Gradient Boosting Machines).
    • Hybrid models that combine raw images with clinical data are also used for more accurate predictions.
    • Significant flaws exist in current datasets, including imbalance and small size, which limit model robustness and generalization.
    • Labels for classification relate to the actual outcomes of procedures, such as live birth potential.
  • Decision-Making Support: AI is used to design protocols and procedures that help embryologists select the best quality oocytes or perform other daily operations.
    • The aim is to propose actions to maximize development potential, shifting from a success/failure modeling to a course-of-action modeling.
    • Examples include optimizing trigger timing for ovarian stimulation or guiding clinicians through daily oocyte management cycles, providing recommendations on drug dosage or follow-up duration.

Discussion and Future Perspectives

The review emphasizes that defining oocyte quality in usable terms for AI solutions is crucial, necessitating an operational definition based on observable and understandable features that allow for objective assessment of viability with high accuracy. This requires integrating top-down models informed by reproductive medicine hypotheses with data-driven models that reveal the role of specific features. An improved understanding of oocyte quality is expected to impact the entire IVF process, potentially shortening time to pregnancy and reducing the number of embryos needing to be frozen.

Key areas for future research and improvement include:

  • Cytoplasmic Movement: This is considered an ideal candidate for both explanatory theory and prediction of live birth potential, but current time-lapse data often has insufficient sampling rates (minutes instead of seconds) to reliably capture these dynamics. Studies should determine the appropriate temporal resolution for imaging techniques to capture these movements accurately.
  • Digital Replicas/Twins of Oocytes: The creation of physically realistic multi-physics models of oocytes could enable physically informed machine learning and provide insights into hidden mechanisms.
  • Multi-Center Studies and Standardized Protocols: There is a compelling need for multi-center studies to collect diverse datasets and reduce biases, enhancing the robustness and generalizability of AI models. Standardized imaging protocols across clinics are essential to ensure consistent and reliable data for comparing results and generalizing AI models.
  • Open-Source Software and Data Sharing: To ensure transparency and reproducibility, making code and high-quality datasets publicly available (while respecting patient privacy) is crucial for encouraging collaboration and knowledge growth within the research community.
  • Focus on Oocytes (Pre-ICSI): Current research heavily focuses on embryos, but shifting analysis to pre-ICSI oocytes could help overcome issues related to the irreversible ICSI procedure and subsequent embryo selection.

The authors acknowledge that despite promising preliminary results from AI-based techniques, challenges persist. These include the availability of high-quality, diverse, and annotated datasets, the need for standardized imaging protocols and data formats, and the integration of simulation approaches. Ethical concerns, beyond data privacy, such as the potential introduction of biases in the IVF process, also need to be considered.

Conclusion

The article concludes that while the application of AI-based techniques to IVF procedures is still in its early stages, there is ample consensus that it could significantly improve both the process and the quality of life for patients. AI’s potential in IVF extends to various phases, including oocyte segmentation, classification, feature extraction, and decision support systems, providing clinicians with data-driven insights. However, the successful integration of AI into current medical practice necessitates a tight, collaborative effort between AI researchers, embryologists, and medical professionals, to address existing challenges related to data quality, standardization, and fundamental scientific understanding of biological phenomena like cytoplasmic movements. This collaboration is seen as essential for achieving true progress in the field and for ensuring that AI acts as an effective complement to human expertise.

Podcast

Course outline :
Artificial Intelligence in Assisted Reproductive Technology: Enhancing Precision, Efficiency, and Outcomes in Modern Fertility Care

This course aims to provide a comprehensive overview of the role of AI in Assisted Reproductive Technologies (ART), its fundamental principles, current applications across various stages of IVF, and the challenges and future prospects of its integration into reproductive medicine.

Duration: 1 Hour 30 Minutes

  • Understanding Infertility and ART:
    • Infertility is a significant global concern, affecting approximately one in six couples worldwide. In the U.S., it impacts about 8.8% of the population, while in Taiwan, it affects 10-15% of couples.
    • ART encompasses fertility treatments involving the handling of eggs or embryos, such as surgically removing eggs, fertilizing them in a lab, and transferring them back.
    • Despite technological advancements, IVF success rates, particularly live birth rates, typically hover around 30% per cycle, highlighting persistent challenges.
  • Introduction to Artificial Intelligence (AI):
    • AI is a sophisticated technological framework that emulates human cognitive functions, enabling computational systems to learn, comprehend, solve problems, and execute tasks.
    • In medicine, AI enhances interpretation of medical imaging, assists in genomic analysis for personalized treatments, and supports precise clinical decision-making.
  • Why AI in ART?
    • AI is becoming an indispensable tool in reproductive medicine, significantly improving the precision and efficiency of fertility treatments.
    • It addresses challenges like the labor- and time-intensive nature of IVF, coupled with significant inter- and intra-observer variability, by augmenting efficiency, reproducibility, and consistency through automation.
  • Data Input and Processing for AI:
    • Data is essential for training AI models to recognize patterns and improve predictions.
    • Effective AI training in ART requires extensive data on patient demographics, treatment protocols, and medical histories.
    • Advanced imaging techniques, such as time-lapse embryo monitoring, provide critical insights into embryo development.
    • Data quality, standardization, and extraction of relevant information are crucial for refining AI models and enhancing their accuracy in ART.
  • Machine Learning Algorithms: AI in ART leverages various machine learning paradigms:
    • Supervised Learning: Trains algorithms using labeled datasets to make future predictions or classify data accurately. In ART, this is used for embryo classification, sperm analysis, and prediction of treatment outcomes.
    • Unsupervised Learning (UL): Autonomously processes extensive datasets without needing human direction to uncover hidden patterns. In ART, UL improves embryo selection, tailors patient treatments, identifies new factors affecting outcomes, and enhances treatment protocols.
    • Reinforcement Learning (RL): An AI system learns by dynamically interacting with its environment using a trial-and-error method to achieve goals, balancing exploration and exploitation. In ART, RL can significantly improve treatment protocols, especially in hormone therapy for IVF, by analyzing patient responses to tailor plans.
  • Controlled Ovarian Stimulation (COS) (10 minutes):
    • Initial Gonadotrophin Dose and Adjustment: AI plays a pivotal role in automating stimulation progression, calibrating dosage adjustments, and optimizing the initial FSH dosage.
      • Machine learning models have demonstrated superiority over clinicians in predicting optimal initial FSH doses for IVF, more precisely targeting the retrieval of 10-15 eggs.
      • AI can customize starting FSH doses to enhance IVF outcomes and reduce gonadotrophin consumption.
    • Follicle Monitoring and Ultrasound Analysis:
      • AI-aided ultrasound transforms follicular monitoring by facilitating automated detection and measurement of follicle sizes and numbers, reducing time intensity and operator variability.
      • 3D ultrasound with AI integration offers comprehensive volumetric views, enhancing prediction of oocyte maturity and optimizing hCG timing.
      • Innovative portable ultrasound systems enable at-home fertility assessments, enhancing continuous care and patient convenience.
    • Decision-Making for Best Day of Trigger:
      • AI models can optimize the timing of trigger injections by integrating various data, such as follicle size and estradiol levels, surpassing traditional physician judgment.
      • This improves retrieval rates of mature oocytes and blastocysts, potentially boosting IVF success rates. AI can also streamline IVF workflow by identifying optimal monitoring days and forecasting trigger timing.
  • Laboratory Procedures (15 minutes):
    • Oocyte Morphological Assessments:
      • AI and Deep Learning (DL) technologies are transforming oocyte quality assessment by offering objective and consistent analysis of meiotic stages, enabling more accurate prediction of fertilization potential and improved pregnancy outcomes.
      • DNN-based systems can classify human oocytes by meiotic maturity with high accuracy (e.g., 96.4% in validation, 95.7% in testing).
      • CNNs and SVMs have significantly improved the prediction of oocyte developmental potential, achieving over 86% accuracy.
    • Semen Analysis:
      • AI and DL enhance sperm selection for IVF and ICSI by automating sperm detection, comprehensive semen analysis, and assessments of sperm viability and DNA integrity.
      • AI systems trained on detailed datasets accurately identify optimal sperm morphology and classify sperm by motility, improving upon traditional methods.
      • Deep learning models can predict DNA integrity from brightfield images, aiding in the selection of high-integrity sperm for ICSI.
    • Embryo Selection and Ploidy Prediction:
      • AI-based tools offer a standardized, non-invasive approach to prioritizing high-quality embryos for transfer, analyzing morphological features to estimate the likelihood of pregnancy.
      • AI models can accurately predict embryo ploidy from images and videos, enhancing IVF outcomes by enabling better embryo selection (e.g., ERICA achieves 65.3% to 77.4% accuracy in predicting euploidy).
      • Time-lapse incubators support AI-driven embryo selection by providing continuous monitoring of development.
  • Micromanipulation Procedures (5 minutes):
    • AI, with its advanced image-processing capabilities, provides precise guidance for intricate procedures such as ICSI and assisted hatching.
    • Deep learning CNNs have achieved high precision (98.9% for polar body location in ICSI and 99.41% for optimal hatching site in AH).
    • Automated Intracytoplasmic Sperm Injection (ICSIA) robots have demonstrated effectiveness, performing ICSI with minimal human intervention and resulting in successful fertilizations and births.
  • Outcome Prediction (5 minutes):
    • Machine learning refines predictive modeling in IVF by leveraging patient-specific data to enhance the accuracy of clinical outcome predictions, such as live birth and clinical pregnancy rates.
    • Models like the random forest algorithm have shown higher accuracy in predicting clinical pregnancies compared to traditional methods.
    • Hybrid AI models integrating video data with clinical features have enhanced predictions of clinical pregnancy outcomes.
  • A. Patient Safety and Quality Management:
    • AI supports flawless execution of IVF lab procedures to prevent human error, such as gamete loss or mismatched transfers.
    • Electronic witnessing systems using RFID and barcodes effectively reduce sample mix-ups (mismatch rates at 0.251%).
    • CNNs have achieved 100% accuracy in identifying patients based on their embryo images, enhancing tracking accuracy in IVF laboratories.
  • B. Social Egg Freezing:
    • AI can accurately predict the number of oocytes required for successful social egg freezing or fertility preservation based on clinical values like age, AFC, and AMH levels.
    • AI models provide precise age-related guidelines, for example, suggesting women under 35 freeze 15 oocytes for a 70% chance of live birth, and 25 for a 95% chance.
  • C. Automation in ART:
    • Automation aims to address the shortage of skilled embryologists and clinicians by reducing reliance on traditional tasks, allowing practitioners to focus on higher-level responsibilities.
    • AI-powered automation improves efficiency, accuracy, and consistency, reducing variability and potentially making fertility care more affordable and accessible.
    • Ongoing advancements include automated sperm preparation, oocyte retrieval, denudation, dish preparation, and cryopreservation.
  • D. Challenges and Limitations of AI in ART:
    • High implementation costs for hardware and software, and the need for extensive staff training.
    • Regulatory hurdles and ensuring patient safety and data privacy remain complex.
    • Data quality concerns and the difficulty in obtaining large, diverse datasets due to the sensitive nature of reproductive data.
    • Bias in AI models if trained on non-representative datasets, potentially leading to disparities in IVF outcomes.
    • Erosion of expertise if clinicians rely too heavily on AI.
    • Lack of transparency (“black box” issue) in many AI systems can erode trust and lead to misinterpretation.
    • Ethical concerns regarding embryo selection and genetic predictions, raising questions about “designer babies”.
    • Environmental impact due to energy-intensive computational processes and expanding data centers.
    • Slow adoption rates in IVF compared to other medical sectors, as AI needs to demonstrate significant advantages and seamless integration.
  • AI is a transformative advancement in ART, enhancing accuracy, efficiency, and clinical decision-making.
  • It serves as an ally to medical professionals, complementing human expertise rather than replacing it, fostering a partnership where AI’s efficiency supports clinicians’ judgment.
  • For widespread adoption, AI solutions must demonstrate consistent, tangible benefits, inspire trust through transparency, adhere to clear regulations, and integrate intuitively into current IVF practices.
  • Future research should focus on incorporating genetic data, exploring lifestyle factors, and validating algorithms across diverse global populations to universalize predictive accuracy.
  • The future of AI in ART hinges on balancing technological innovation with the needs, expectations, and expertise of the professionals it aims to support.

Slides for the Powerpoint Presentation

Slide 1: AI in Assisted Reproductive Technology: Enhancing Fertility Outcomes

 

  • Course Duration: 1 Hour 30 Minutes
  • Course Objective: Provide a comprehensive overview of AI’s role in ART, its principles, applications across IVF stages, and the challenges and future prospects of its integration into reproductive medicine.
  • Presented by: [Your Name/Organization]
  • Date: [Date]

 

Slide 2: The Landscape of Infertility and Assisted Reproductive Technology (ART)

 

  • Infertility affects approximately one in six couples worldwide, highlighting a significant global health concern.
  • In the United States, about 8.8% of the population is impacted by infertility, while in Taiwan, this figure ranges from 10-15% of couples.
  • ART encompasses fertility treatments that involve the handling of eggs or embryos, such as surgical egg retrieval, laboratory fertilization, and embryo transfer.
  • Despite significant technological advancements since 1978, the live birth rate in IVF typically hovers around 30% per cycle, indicating persistent challenges and room for improvement.
  • The average duration for diagnosing infertility in Taiwan stretches to 2.9 years, notably surpassing the WHO’s one-year benchmark.

 

Slide 3: Demystifying Artificial Intelligence: A Foundation

 

  • AI is a sophisticated technological framework that emulates human cognitive functions, enabling computational systems to learn, comprehend, solve problems, and execute tasks.
  • In medicine, AI enhances the interpretation of medical imaging, assists in genomic analysis for personalized treatments, and plays a crucial role in precise clinical decision-making.
  • The concept of automation, a key benefit leveraged by AI, involves mechanizing tasks traditionally performed by humans through computerized systems.
  • AI also accelerates drug discovery, eases patient monitoring, and enhances surgical procedures through robotic assistance.
  • AI’s influence in healthcare is evident through rapid image interpretation, workflow efficiency, and patient empowerment via apps and wearables.

 

Slide 4: Why Artificial Intelligence is Indispensable in ART

 

  • AI is becoming an indispensable tool in reproductive medicine, significantly improving the precision and efficiency of fertility treatments.
  • IVF is a sophisticated, multistage procedure that is labor- and time-intensive, coupled with significant inter- and intra-observer variability.
  • AI addresses these challenges by augmenting efficiency, reproducibility, and consistency through automation.
  • By automating routine and time-consuming steps, AI has the potential to alleviate the workload of physicians and embryologists.
  • Through continuous machine learning, AI guarantees optimal practices and results while reducing the likelihood of errors.
  • The integration of AI aims to help address the shortage of skilled embryologists and clinicians, reducing reliance on traditional tasks.

 

Slide 5: The Bedrock of AI in ART: Data Input and Processing

 

  • Data is essential for training AI models to recognize patterns and improve predictions over time.
  • Effective AI training in ART requires extensive data on patient demographics, treatment protocols, and medical histories.
  • Advanced imaging techniques, such as time-lapse embryo monitoring, provide critical insights into embryo development.
  • Data quality, standardization, and extracting relevant information are crucial for refining AI models and enhancing their accuracy in ART.
  • However, obtaining large, diverse datasets is challenging due to the sensitive and variable nature of reproductive data, and data security remains a concern.
  • Current data sources are mainly still images, time-lapses, and clinical data, but image capture often lacks standardization across clinics.

Slide 6: The Engine of AI: Machine Learning Paradigms

 

  • Machine learning algorithms enable computers to learn from data and make predictions or decisions.
  • They encompass a wide range of techniques, from supervised learning (trained on labeled data) to unsupervised learning (uncovers hidden patterns in unlabeled data).
  • Reinforcement learning allows an AI system to learn by dynamically interacting with its environment through trial-and-error.
  • The choice of machine learning approach (supervised, unsupervised, or reinforcement learning) should be driven by the task at hand, the nature of the data, and available resources.
  • Classical machine learning models used in medicine include linear regression, logistic regression, decision trees, random forests, and support vector machines (SVM).

 

Slide 7: Supervised Learning: Predicting Outcomes with Labeled Data

 

  • Supervised learning involves training algorithms using labeled datasets, where each data point is associated with a target variable.
  • The quality and quantity of training data significantly influence the accuracy of these predictions.
  • In ART, this approach is frequently employed for tasks such as embryo classification, sperm analysis, and the prediction of treatment outcomes.
  • For example, Deep Learning (DL) or SVM methods have shown high sensitivity and specificity in selecting spermatozoa with high-quality morphology.
  • An Artificial Neural Network (ANN) trained with 12 features, including woman’s age and endometrial thickness, showed 76.7% sensitivity and 73.4% specificity in predicting live births.

 

Slide 8: Advanced Learning in ART: Uncovering Patterns and Optimizing Protocols

 

  • Unsupervised Learning (UL) autonomously processes extensive datasets without requiring human direction, uncovering hidden patterns and structures.
  • In ART, UL is invaluable for improving embryo selection, tailoring patient treatments, identifying new factors affecting ART outcomes, and enhancing treatment protocols.
  • Reinforcement Learning (RL) allows an AI system to learn by dynamically interacting with its environment, using a trial-and-error method to achieve goals.
  • In ART, RL can significantly improve treatment protocols, particularly in hormone therapy for IVF, by analyzing patient responses to tailor plans more personally, potentially increasing success rates.

 

Slide 9: AI’s Journey Through IVF: A Multistage Enhancement

 

  • AI provides comprehensive assistance in various ART treatments and at different stages of the IVF process.
  • Its applications span ovarian stimulation, diverse aspects in the IVF laboratory, embryo transfer, and patient safety and quality management.
  • Recent research in fertility journals showcases AI’s transformative role, particularly in enhancing embryo selection and optimizing other IVF-related procedures.
  • AI’s multifaceted impact extends to patient care, treatment efficiency, and success rates.
  • The objective of AI in IVF is to enhance the efficacy of treatments and the clinical decision-making process.

 

Slide 10: Optimizing Ovarian Stimulation: AI’s Role in Dosage and Adjustment

 

  • AI plays a pivotal role in automating stimulation progression, calibrating dosage adjustments, and optimizing the initial FSH dosage.
  • Machine learning models have demonstrated superiority over clinicians in predicting optimal initial FSH doses for IVF, more precisely targeting the retrieval of 10-15 eggs.
  • AI can customize starting FSH doses to enhance IVF outcomes and reduce gonadotrophin consumption.
  • One model predicted the collection of an additional 1.5 mature MII oocytes on average for dose-responsive patients.
  • An algorithm assessed in 2603 IVF cycles showed high decision-making accuracy (0.92 for continue/stop treatment, 0.96 for trigger/cancel) for ovarian stimulation management.

Slide 11: AI-Aided Follicle Monitoring and Ultrasound Analysis

 

  • AI-aided ultrasound transforms follicular monitoring by facilitating automated detection and measurement of follicle sizes and numbers, reducing time intensity and operator variability.
  • 3D ultrasound with AI integration offers comprehensive volumetric views, enhancing prediction of oocyte maturity and optimizing hCG timing.
  • Studies have identified superior volume cut-offs for oocyte retrieval and hCG administration using deep learning models on 3D ultrasound data.
  • Innovative portable ultrasound systems enable at-home fertility assessments, enhancing continuous care and patient convenience with a 96% success rate.
  • AI models like CR-Unet and HaTU-Net achieve high accuracy in segmenting ovarian and follicular structures on transvaginal ultrasound images.

 

Slide 12: Precision Triggering: AI for Optimal Oocyte Retrieval Timing

 

  • AI models can optimize the timing of trigger injections by integrating various data, such as follicle size and estradiol levels, often surpassing traditional physician judgment.
  • A machine learning algorithm in 7866 ICSI cycles increased the average number of fertilized oocytes by 1.430 and usable blastocysts by 0.577 per IVF cycle.
  • AI improves retrieval rates of mature oocytes and blastocysts, potentially boosting IVF success rates.
  • AI can streamline IVF workflow by identifying optimal monitoring days and forecasting trigger timing, potentially reducing patient visits.
  • One study revealed that scheduling follicular tracking scans on Days 5-7 of stimulation is key for optimal IVF planning, allowing prediction of trigger day and OHSS risk.

 

Slide 13: AI in the Lab: Revolutionizing Oocyte Assessment

 

  • AI and Deep Learning (DL) technologies are transforming oocyte quality assessment by offering objective and consistent analysis of meiotic stages.
  • This enables more accurate prediction of fertilization potential and improved pregnancy outcomes by evaluating stages like metaphase II (MII), metaphase I (MI), and germinal vesicle (GV) non-invasively.
  • A DNN-based system can classify human oocytes by meiotic maturity with high accuracy, achieving 96.4% in validation and 95.7% in testing.
  • CNNs and SVMs have significantly improved the prediction of oocyte developmental potential, achieving over 86% accuracy.
  • AI algorithms like Violet (Future Family) boast up to 90% accuracy in predicting fertilization outcomes from a single oocyte image.

 

Slide 14: AI in the Lab: Enhancing Semen Analysis and Sperm Selection

 

  • AI and DL enhance sperm selection for IVF and ICSI by automating sperm detection, comprehensive semen analysis, and assessments of sperm viability and DNA integrity.
  • AI systems trained on detailed datasets accurately identify optimal sperm morphology and classify sperm by motility, improving upon traditional methods.
  • A new deep learning algorithm accurately identifies morphological abnormalities in sperm acrosome and vacuole regions, aiding ICSI selection and operating in real-time.
  • AI-powered tools excel at classifying sperm morphology, identifying subtle defects with high accuracy, with automated classifiers achieving up to 94% accuracy rates.
  • Computer-aided sperm analyzers (CASA) are useful tools for rapid analysis, reducing interoperator variability and providing more accurate measures of sperm motility and kinematics.

 

Slide 15: AI in the Lab: Advanced Sperm Motility and DNA Integrity Analysis

 

  • The shift to computer-assisted sperm analysis (CASA) has standardized and accelerated motility evaluations globally.
  • AI, specifically through convolutional neural networks, allows for precise and quick analysis of sperm trajectories, greatly improving the accuracy of semen analysis.
  • Deep learning models effectively and efficiently predict sperm motility from video data alone, with results delivered in under 5 minutes.
  • The Bemaner system uses AI and cloud technology to enable home-based sperm motility analysis, validated by medical experts.
  • A new deep learning model can assess DNA integrity from brightfield images, correlating moderately at 0.43 and selecting high-integrity sperm in under 10 ms each, addressing the invasiveness of traditional methods.

 

Slide 16: AI in the Lab: Innovating Embryo Selection and Ploidy Prediction

 

  • AI-based tools offer a standardized, non-invasive approach to prioritizing high-quality embryos for transfer, analyzing morphological features to estimate the likelihood of pregnancy.
  • AI models can accurately predict embryo ploidy from images and videos, enhancing IVF outcomes by enabling better embryo selection.
  • The ERICA (Embryo Ranking Intelligent Classification Algorithm) system, trained on Day 5 embryo images, predicts ploidy potential and probability of pregnancy, potentially surpassing human embryologists in accuracy.
  • Time-lapse incubators support AI-driven embryo selection by providing continuous, non-disruptive monitoring of embryo development.
  • AI models like random forest have shown higher accuracy in predicting clinical pregnancies compared to traditional methods.
  • While AI reduces assessment time and standardizes evaluations, current evidence does not conclusively show better clinical outcomes compared to traditional methods.

 

Slide 17: AI in Micromanipulation: Precision in IVF Laboratory Tasks

 

  • AI, with its advanced image-processing capabilities, provides precise guidance for intricate procedures such as ICSI (Intracytoplasmic Sperm Injection) and assisted hatching (AH).
  • Deep learning CNNs have achieved high precision in identifying critical micromanipulation sites: 98.9% accuracy for polar body location in ICSI and 99.41% for optimal hatching site in AH.
  • Automated Intracytoplasmic Sperm Injection (ICSIA) robots have demonstrated effectiveness, performing ICSI with minimal human intervention and resulting in successful fertilizations and births.
  • The integration of microscopic spindle visualization into ICSIA seeks to enhance the accuracy and safety of IVF procedures, potentially improving fertilization rates and pregnancy outcomes.
  • This technology could expand micromanipulation applications beyond ICSI and assisted hatching to include trophectoderm biopsy.

 

Slide 18: Forecasting Success: AI in IVF Outcome Prediction

 

  • Machine learning refines predictive modeling in IVF by leveraging patient-specific data to enhance the accuracy of clinical outcome predictions, such as live birth and clinical pregnancy rates.
  • Models like the random forest algorithm have shown higher accuracy in predicting clinical pregnancies compared to logistic regression (AUC of 0.7208 vs 0.6766).
  • Hybrid AI models integrating video data with clinical features have enhanced predictions of clinical pregnancy outcomes (AUC of 0.727 vs 0.684 for video-only).
  • Key predictive factors identified include embryo morphokinetics, oocyte age, gonadotrophin dosage, and endometrium thickness.
  • Dynamic, data-driven models can accurately predict live birth outcomes in IVF, enabling real-time updates and decision support throughout the process.

 

Slide 19: Ensuring Safety and Quality: AI in ART Laboratories

 

  • AI supports flawless execution of IVF lab procedures to prevent human error, such as gamete loss or mismatched transfers.
  • Electronic witnessing systems (EWS) using RFID and barcodes effectively reduce sample mix-ups, with mismatch rates as low as 0.251%.
  • CNNs have achieved 100% accuracy in identifying patients based on their embryo images on days 3 and 5, enhancing tracking accuracy in IVF laboratories.
  • HFEA guidelines emphasize strict witnessing protocols in ART, advocating for electronic tracking systems and requiring contingency plans for electronic system failures.
  • EWS has tracked over 849,650 points in 109,655 IVF, ICSI, FET, and IUI cycles over a decade.

 

Slide 20: Empowering Choices: AI and Social Egg Freezing

 

  • Social egg freezing allows women to delay childbearing for non-medical reasons, offering flexibility to pursue personal, professional, or educational goals.
  • AI can accurately predict the number of oocytes required for successful social egg freezing or fertility preservation based on clinical values like age, AFC, and AMH levels.
  • AI models provide precise age-related guidelines, for example, suggesting women under 35 freeze 15 oocytes for a 70% chance of live birth, and 25 for a 95% chance.
  • For women at 37, 20 oocytes are needed for a 75% chance of live birth, and at 42, 61 oocytes are required for the same success rate.
  • These insights support personalized fertility strategies, enhancing patient-doctor discussions and boosting confidence in medical choices.

Slide 21: The Automated Future: Transforming ART Workflows

 

  • Automation aims to address the shortage of skilled embryologists and clinicians by reducing reliance on traditional tasks, allowing practitioners to focus on higher-level responsibilities.
  • AI-powered automation improves efficiency, accuracy, and consistency, reducing variability and potentially making fertility care more affordable and accessible.
  • Ongoing advancements include automated sperm preparation, oocyte retrieval, denudation, dish preparation, and cryopreservation.
  • Automated ICSI robots have successfully injected human oocytes, leading to births from partially automated procedures.
  • AI-powered robotic systems can perform remote automated ICSI with comparable outcomes to manual methods, paving the way for standardized procedures.

 

Slide 22: Navigating Hurdles: Implementation and Data Challenges for AI in ART

 

  • High implementation costs for hardware and software, and the need for extensive staff training, are significant practical challenges.
  • Integrating AI into clinical settings requires substantial investment in both technology and personnel development.
  • AI models depend on large, diverse datasets for training, which can be difficult to obtain due to the sensitive and variable nature of reproductive data.
  • Data quality concerns and the difficulty in obtaining large, diverse datasets are major limitations, as most studies rely on small, single-clinic datasets.
  • The lack of standardization in image capture (different microscopes, lenses, cameras) introduces bias and complicates general-purpose algorithm implementation.

 

Slide 23: Broader Concerns: Ethics, Regulation, and Sustainability of AI in ART

 

  • Bias in AI models if trained on non-representative datasets can perpetuate disparities and potentially impact IVF outcomes for specific populations.
  • Erosion of expertise is a concern if clinicians rely too heavily on AI, potentially degrading their decision-making skills.
  • Lack of transparency (“black box” issue) in many AI systems can erode trust and lead to misinterpretation if outputs are blindly adopted.
  • Ethical concerns regarding embryo selection and genetic predictions raise questions about “designer babies” and can lead to societal pushback.
  • Regulatory hurdles are complex, with frameworks often lagging behind AI’s rapid evolution and global harmonization remaining limited.
  • Environmental impact due to energy-intensive computational processes and expanding data centers is a growing concern, with US data center electricity consumption projected to rise significantly.

 

Slide 24: A Transformative Ally in Reproductive Medicine

 

  • AI is a transformative advancement in ART, enhancing accuracy, efficiency, and clinical decision-making across various IVF stages.
  • It serves as an ally to medical professionals, complementing human expertise rather than replacing it, fostering a partnership where AI’s efficiency supports clinicians’ judgment.
  • AI optimizes protocols from ultrasound imaging for follicle assessment to precise gonadotropin dosing and identifying ideal embryo transfer sites.
  • In the IVF lab, AI excels in micromanipulation techniques, enhancing the precision of procedures such as ICSI and assisted hatching.
  • AI aids clinicians in selecting the optimal embryos for transfer or cryopreservation.
  • The future of AI in ART hinges on balancing technological innovation with the needs, expectations, and expertise of the professionals it aims to support.

 

Slide 25: The Future of AI in ART: Prospects and Priorities

 

  • For widespread adoption, AI solutions must demonstrate consistent, tangible benefits, inspire trust through transparency, adhere to clear regulations, and integrate intuitively into current IVF practices.
  • Future research should focus on incorporating genetic data, exploring lifestyle factors, and validating algorithms across diverse global populations to universalize predictive accuracy.
  • The precision of ART could further increase through personalization of stimulation protocols and more accurate selection of gametes and embryos.
  • The creation of properly designed multicenter studies is crucial to validate machine learning models on large populations, overcoming current limitations of small sample sizes.
  • Clinicians must work in an integrated manner with machine learning, without blindly relying on it, to avoid professional disqualification and loss of trust.
  • The opportunity for AI to not only enhance IVF outcomes but also redefine the role of embryologists, enabling them to focus on higher-level, value-driven tasks, remains significant.