Search Strategy

A comprehensive literature search was conducted between 28th June 2023 and 5th July 2023 in PubMed and Elsevier Library using the keywords ‘Deep Learning’ OR ‘Machine Learning’ OR ‘Artificial Intelligence’ OR ‘Computer Vision’ AND ‘Blastocyst’ OR ‘Embryo.’ An additional search was performed on January 15, 2025, to update the existing literature.

Study Selection

The initial search retrieved 384 studies. The inclusion criteria for the review were articles published in the last 10 years that focused on the most relevant studies, resulting in 362 studies. After excluding 32 duplicates, the remaining studies were screened to include only studies published in English and those based on human blastocysts, leaving 186 studies to be screened.

Four other eligible studies were identified through thorough forward and backwards citation mining of the selected relevant literature by employing the Open Knowledge Maps Tool on 29th August 2024.

Excluded Studies

Articles related to the use of AI models in the assessment of cleavage-stage embryos (day 3 embryos) and oocyte morphology were excluded. Additionally, generic narratives on the use of AI in IVF labs and gamete assessments were excluded. Studies that used AI for the association of clinical parameters and live births with proteomic, transcriptomic, and radiomic models were also excluded. Furthermore, we excluded studies that solely aimed to assess the efficacy of existing AI models, such as the Intelligent Data Analysis (iDA) and Known Implantation Data (KID) scores, on specific datasets annotated by the authors. These studies did not involve training the AI models but rather focused on applying pre-trained models to datasets with clinical or morphological annotations.

Primary selection was performed based on title and abstract screening using the PRISMA guidelines, which resulted in 26 eligible studies.

Data Extraction

Data extraction was independently performed by two reviewers Doel Bose and Sayali Kandari (DB and SK). A consensus was reached through discussion, and a third reviewer Sandeep Kumar Verma (SKV) resolved any persistent disagreements.

A standardised spreadsheet was used for meticulous data extraction, facilitating the systematic collection of essential information from each included study. The key parameters were as follows:

• Dataset size: Total number of samples/images in each dataset.

• Span of data collection: Span of data collection in years.

• Dataset source clinics: Clinic(s) or providers of the dataset.

• Image capture technique: Static image or time-lapse video.

• Image resolution: Pixel dimensions or image quality within the captured dataset.

• Image/video capture device: Camera/device vendor and type.

• Metadata with image: Maternal age and clinical parameters used to train the AI model.

• Dataset outcome or ground truth or endpoint: Was the outcome measured in terms of live birth or cardiac activity? If cardiac activity was measured, when was it measured, 6 weeks | 8 weeks?

• Day of image capture: blastocyst image capture day—D5 or D6, or D7

• Dataset capture timing: When was the image captured in terms of hpi.

• Number of focal planes used for image/video capture: The number of focal planes utilised in capturing images or videos of embryos intended for AI model training.

• Dataset class distribution: The distribution of images across different classes or categories, that is, the percentage of images with a negative outcome and the percentage of images with a positive outcome.

• Type of embryo transfer: Whether images are related to fresh embryo transfer, frozen embryo transfer (FET), or a combination of both strategies.

• Public availability of data: Whether the datasets were publicly available for use by other researchers.

Quality Assessment

Two authors independently reviewed the abstracts and excluded those that did not meet the inclusion criteria. The selected papers were assigned to both reviewers [Doel Bose (DB) and Sayali Kandari (SK)], with conflicts resolved by a third author Hemant Deshmukh (HD) through majority voting. Data extraction and quality assessment were conducted independently by two authors, and disagreements were resolved by a third author. Since the majority of studies in our review did not strictly qualify as medical interventions, a more tailored approach was chosen to assess the risk of bias in datasets using predefined criteria for comparison using spreadsheets.

Variability in Dataset Size

In our comprehensive review of 26 studies, each study was thoroughly scrutinised to ascertain the actual number of images used to train the AI model, as shown in Table 1.^{[15,17,18-35]} Many papers presented an inflated number in the dataset size, but later in the ‘Material and Methods’ section, many studies mentioned the exclusion of images, as they were of poor quality, or mentioned data augmentation by flipping and rotations. Thus, only filtered images that were fed to the AI model during training or validation without data augmentation were considered for the dataset size.

A wide range of dataset sizes was observed, ranging from a minimum of 160 images to a maximum of 171,239 images, indicating the diversity of the dataset size, as depicted in Figure 2.^[17,18] It is important to note that two studies utilised a dataset size of less than 500 images, whereas two others utilised a dataset size of more than 100,000 images.^{[17,18,36,37]} It is intuitively understood that a more extensive dataset contributes to the enhanced training and performance of AI models in blastocyst assessment.

Close

Figure 2:

Dataset size used in studies. Two studies used less than 500 images to train the AI model. Nine studies utilised as many as 1000 images to train the AI model, while 2 studies used greater than 1,00,000 images for training the AI model.

Export to PPT

Dataset Demographics and Source

Most studies (15 out of 26) have focused their AI models on data from a single clinic, potentially limiting adaptability, as shown in Table 1.^{[17,18,20–24,28,30,32,37–39]} Notable exceptions include a study using a diverse multinational dataset (USA, India, Spain and Malaysia) and another incorporating data from 18 clinics across five countries (Denmark, UK, Australia, Japan and Ireland), showing broader geographical coverage as shown in Figure 3.^[31,37]

Close

Figure 3:

Number of clinics involved in data collection. Of the 26 studies included in the systematic review, 15 collected data from a single clinic. Only three studies collected diverse data that spanned more than one country.

Export to PPT

Variability in Image Capture Mechanism and Resolution

A diverse range of image resolutions (100 pixels × 100 pixels to 1024 pixels × 768 pixels) reflects various equipment choices as shown in Table 2. Notably, 13 studies employed only time-lapse devices, while eight solely utilised optical light microscopes, illustrating technological diversity.^{[18,25,29,31,36,38,39]} The number of focal planes in the time-lapse video ranged from 7 to 11, introducing complexity to the data collection method as shown in Figure 4.

Close

Figure 4:

Focal planes used to capture images in datasets. Out of 26 studies, 17 collected images using a single focal plane, i.e., static images, while a few studies relied on capturing images of the same blastocyst at varying focal planes ranging from 2 to 11 focal planes.

Export to PPT

In the majority of studies (12 of 26), a single data capture mechanism was used exclusively. Two studies did not specify their image capture technique.^[17,24] Only two studies utilised more than one image-capture mechanism. A study by Diakiw, S. M.,^[27] predominantly used static images from optical light microscopes for training and validation, reserving static frames from time-lapse videos specifically for testing only, as shown in Figure 5. Loewke, K.^[15] incorporated low-resolution images (56 pixels × 56 pixels) from a stereo zoom microscope as part of their training data, while also using static frames from time-lapse videos exclusively to test their model.

Close

Figure 5:

Dataset image capture source. Out of 26 studies included in the systematic review, 14 collected images from the time-lapse system, while eight studies relied solely on static images collected from an optical light microscope.

Export to PPT

Dataset Image Capture Camera Model

A notable observation made while evaluating the camera model of the dataset across studies is shown in Figure 6. Among the 26 studies, more than half (17 out of 26) used a single type of device to capture images, either a single type of optical light microscope or a single type of time-lapse incubator fitted with a single type of camera.^{[18–25,28,30,32,33,35,36,39,40]} Exceptions to these findings are two studies that included six and eight camera models, respectively, for collecting input datasets, as shown in Figure 6.^[15,27]

Close

Figure 6:

Types of image capture mechanisms used in the collection of image datasets. Thirteen studies used images from a single type of optical system while seven did not specifically mention the make or model of the optical system and camera. Two studies utilized six and eight types of image capture mechanisms, as shown.

Export to PPT

Dataset Image Capture Timing

The timing of blastocyst image capture presents a significant discrepancy among studies. Thirteen studies lacked specific mention of the image or video capture timing on day 5 or day 6.^{[15,26–28,31,33,36,38–40,41]} Among those providing the exact timing, seven studies utilise images captured at the exact time range between 110 and 115 hpi, as shown in Table 3.^{[17,19,20–23,25]} Additionally, six studies incorporated a wider range of timing, capturing images anywhere between 104 and 140 hours, i.e., day 5 to day 6.^{[18,24,28,30,32,33,37]}

Only 7 out of 26 studies used a fixed time frame on day 5 with respect to hpi to capture blastocyst images, as shown in Figure 7. One study utilised a variable timeframe for image capture, anywhere between days 5, 6, or 7, which raises questions about the quality of the dataset.^[15]

Close

Figure 7:

Day and timing of image capture of datasets. Only 7 out of 26 studies use a fixed time frame with respect to hours post insemination (hpi) to capture a blastocyst image.

Export to PPT

Dataset Endpoints or Ground Truth

In the majority of studies included in this systematic review, the correlation between the dataset endpoint and the intended objective of blastocyst assessment was notable, as shown in Table 4. An exception worth noting was the study which focused on ranking embryos based on their implantation potential.^[29] Here, the premature endpoint involved measuring βHCG at day 7 post-embryo transfer. It is worth noting that a more optimal approach might have been assessing βHCG levels at a later date or monitoring foetal heartbeat at 6–8 weeks for a more accurate evaluation. The bar chart, as shown in Figure 8, represents the endpoints, or ground truth, opted in various studies included within this systematic review.

Close

Figure 8:

Ground truth or end points of studies. Only one study used serum βHCG as end point while nine studies tracked patients till live birth. βHCG: Beta human chorionic gonadotropin, PGT-A: Preimplantation genetic testing for aneuploidy, FHB: Foetal heartbeat.

Export to PPT

Another noteworthy study in the review centres on blastocyst assessment aimed at associating a probability score for live birth. In the study by Miyagi Y,^[17] the endpoint was defined by measuring live births or tracking aneuploid abortions verified by CVS, providing a robust method as compared to relying solely on negative βHCG or missing cardiac activity.^[17]

Dataset Cleansing – Removing Confounders

Most of the studies associated with blastocyst assessment aimed for either viability or predictive score. These studies did not consider filtering the datasets to remove confounding factors like uterine abnormalities or receptivity that could influence the clinical endpoint of foetal heartbeat at 6–8 weeks.

Only one study excluded confounders by only considering aneuploid miscarriages as the negative class, eliminating various confounding factors related to uterine abnormalities.^[17]

Dataset Sampling Strategy and Class Imbalance

Considerable variation in dataset class distribution was observed among the included studies, as detailed in Table 4. A realistic distribution of close to 40% good embryos and 60% poor embryos was commonly employed in 9 out of 26 studies, as shown in Figure 9, while seven studies approached an even 50%–50% split.

Close

Figure 9:

Dataset class distribution—percentage of negative class vs. positive class. Notably, five studies did not have a clear mention of the class distribution. Seven adopted a class distribution of 50%–50% between negative and positive classes, while another nine opted for a distribution closer to reality, with 60% negative and 40% positive classes.

Export to PPT

However, one study employed an unusually imbalanced class distribution, opting for an 8% positive and 92% negative split.^[26] This emphasises the lack of standardised common practices in defining the composition of training datasets and their sampling strategies.

Embryo Transfer Strategy: Fresh vs. Frozen Transfer

Including data from both fresh and FETs in seven studies introduces a confounder to the predictive model, as shown in Table 5.^{[15,22,26–29,37]} It is widely acknowledged that fresh embryo transfers tend to have a lower pregnancy rate than FET cycles, primarily due to compromised uterine receptivity.^[42] Studies focused on predicting embryo viability and euploidy logically opted for FET as their preferred embryo transfer strategy.^{[19,27,30,31]} This choice aligns with the nature of these studies, which relied on predicting euploidy outcomes, where the endpoint was the PGT-A result.

Five studies do not clearly mention the embryo transfer strategy, as shown in Figure 10. Since the endpoints of these studies were the classification or grading of blastocysts, the embryo transfer strategy was not applicable.^{[18,21,20,22,25]}

Close

Figure 10:

Embryo transfer strategy employed in datasets. Seven studies included data from both fresh and frozen transfers, while five failed to mention the embryo transfer strategy.

Export to PPT

Dataset Inclusion and Exclusion Criteria

In a study by Loewke, we recorded the inclusion of discarded or mislabelled blastocysts in the negative outcome class.^[15] However, assuming negative outcomes for all discarded blastocysts might be flawed, as supernumerary blastocysts are often discarded without known outcomes, especially when patients withhold consent for freezing.

Four studies have failed to clarify the embryo transfer strategy, i.e., fresh or FET, employed in the datasets collected, further impacting the IVF cycle outcome.^{[17,24,36,38]} Including such data has the potential to influence model training and overall performance.

Achieving a balance between inclusion and exclusion criteria is a pivotal factor in constructing a dataset that accurately represents the intended predictive endpoint.

Inclusion of Metadata in Training AI Model

Five studies that concentrated on assigning a grade to blastocysts or categorising them as good or poor consciously chose not to incorporate any metadata, as shown in Figure 11.^{[18,21,22,25,32]} This decision aligns with the nature of their task, in which the role of metadata may not be considered significant.

Close

Figure 11:

Different types of metadata used along with the blastocyst image. Out of 13 studies that used metadata, two used only maternal age. Notably, three studies used other clinical parameters apart from maternal age. ICM: Inner Cell Mass, TE: Trophectoderm.

Export to PPT

Surprisingly, few studies that specifically target predicting embryo viability and implantation potential have omitted the use of patient metadata altogether.^{[17,26,31,36,37]} However, a noteworthy study employed a comprehensive approach to train the AI model by incorporating a range of clinical parameters such as maternal age, number of embryo transfers, anti-Müllerian hormone concentration, day-3 blastomere number, grade on day 3, embryo cryopreservation day, ICM, TE, average diameter, and body mass index (BMI), etc. over and above the morphological parameters of the embryo as shown in Figure 11.^[24] A more recent model also integrated maternal age with pseudo-features derived from Istanbul grading to enhance predictive performance.^[34] Notably, automated extraction of morphometric parameters has demonstrated a positive association with implantation potential.^[35]

Another recent study employed a deep learning pipeline using day-5 embryo images along with maternal age and pseudo-features derived from Istanbul grading scores (stage, ICM, and TE), combining them with self-supervised and Generative Adversarial network (GAN)-augmented models to enhance viability prediction. This approach reflects a growing trend toward hybrid models that merge visual and clinical data to improve predictive accuracy.^[34]

Additionally, seven studies that utilise time-lapse images and videos also integrated embryo morphokinetics data into their model training, highlighting the importance that these parameters provide, as shown in Figure 11, recognising the valuable insights these parameters provide. This diversity in metadata utilisation highlights the need for standardised practices and considerations in optimising models for predicting implantation potential.

Among studies using metadata for AI training, three studies utilised maternal age with other clinical data, as shown in Figure 11.^[29,38,39] Four studies relied on morphokinetic data.^{[22,23,28,37]} This diverse approach showcases ongoing experimentation to boost AI models for predicting embryo outcomes, logically emphasising that the use of metadata should enhance the model accuracy.

Public Availability of Data

Of the 26 studies reviewed, a significant portion lacked transparency regarding the dataset availability. Specifically, 14 studies did not explicitly mention whether their datasets were accessible to other researchers. Among the remaining 12, seven studies indicated that anonymised datasets can be obtained upon request, reflecting a willingness to share data while ensuring participant privacy.^{[17,22,30,37,39,40]} However, four studies explicitly denied public access to their embryo images and patient data, citing ethical and privacy reasons and limiting access to collaborators.^{[19,20,27,33]} This disparity in data availability highlights the ongoing challenges of balancing data sharing with ethical considerations in reproductive research.

A consolidated overview of dataset characteristics, including but not limited to study type, dataset size, span of data collection, image resolution, timing of image capture, metadata inclusion, and clinical endpoints, is presented in Table 6, reflecting the wide heterogeneity observed across the 26 included studies.

Comparative summary of dataset characteristics extracted from 26 studies evaluating AI-based blastocyst assessment. The table includes parameters such as dataset size, span of data collection, number of clinics involved in data collection, type of study, public data availability, resolution of image, type of input to the AI model, number of focal planes, types of camera used, timing of image capture, day of image capture, embryo transfer strategy, metadata inclusion in AI model training, primary aim of blastocyst assessment, and the endpoint or ground truth used. This table is provided as a separate Excel file due to formatting constraints and is submitted as Table 6.

Importance of Dataset Size

The impact of the sample size on the performance of AI models for blastocyst assessment is significant and multifaceted. Although a larger sample size intuitively correlates with improved model performance, the quality and distribution of data within the dataset are of paramount importance.^[43] Studies have shown that satisfactory model performance can be achieved even with a smaller sample size that contains high-quality data and is representative of the problem domain.^[44] However, it is essential to recognise that there exists a point of diminishing returns where the effect sizes and ML accuracies plateau after reaching a certain sample size threshold.^[45] This suggests that simply increasing the sample size may not always lead to significant improvements in model performance. Therefore, when evaluating sample size adequacy, it is crucial to consider the quantity, quality, and sampling strategy of data to ensure optimal model performance in blastocyst assessment studies, while also acknowledging the resource constraints involved in obtaining and annotating large datasets.

Importance of Diversity in Training Data

Diverse demographic dataset sources are crucial for robust and universal AI models for blastocyst assessment. Training on varied datasets goes beyond regional boundaries, ensuring adaptability to global populations and IVF laboratory settings. AI models benefit from capturing diverse blastocyst morphologies and morphokinetic trends owing to patient ethnicity, enhancing their applicability across different clinical settings worldwide.^[46,47]

These findings set the stage for a broader discussion of the implications of dataset demographics, emphasising the need for standardised approaches in selecting datasets for AI model training.

In addition to data from diverse demographics, incorporating various image capture mechanisms is intuitively important. While training with diverse image-capturing mechanisms may seem important, there are strategies such as domain adaptation and domain generalisation that can be employed to address the issue of domain shift and enable seamless integration of these AI models into real-world clinical settings.

Domain adaptation enables models trained in one setting to perform well in others, whereas domain generalisation aims to maintain model performance across multiple domains without specific adaptation.^[48,49] Personalised federated learning allows for the training of a shared model with data from multiple clinics, which can then be tailored to individual practices.^[50] These strategies ensure that AI models are versatile and customisable in various clinical settings.

Importance of Using Multiple Focal Planes for Image Capture

Although it may seem intuitive that AI models for embryo assessment could benefit from using multiple focal planes to better visualise key characteristics, such as the ICM and TE, the current research does not provide a solid basis for this claim. In a recent abstract that explored the performance of an AI model trained on static images using a single focal depth, there was no statistically significant difference in model performance when applied to multifocal time-lapse images with different focal planes.^[50,51] Another abstract mentioned that techniques like ensemble learning and test-time augmentation can also be used to reduce the sensitivity of AI models to different focal planes while maintaining model performance.^[52]

However, a recent retrospective study of 2555 day 5 blastocysts showed that training the AI model with segmented images highlighting the ICM and TE improved its predictive performance, with the AUC increasing from 0.716 to 0.741.^[41] The model was also able to focus on clinically relevant regions in 99% of cases, compared to 86% when trained on unsegmented images. This suggests that providing structured input emphasising biologically important regions can enhance the accuracy of AI models in embryo assessment.

Hence, we currently do not have sufficient evidence to suggest that multifocal images will be better for training AI models, and this needs further research with larger sample sizes.

Implications of Wider Range in Image Capture Timing

Blastocysts are highly metabolically active, and even a slight 3–5 hour gap in assessment timing has the potential to significantly impact their grading.^[42,53] None of the reviewed papers, especially those using static blastocyst images, explicitly specified whether they incorporated the timing of image capture as temporal metadata along with the input image to train their AI models tasked with assessment and ranking blastocysts. It is worth noting that the Istanbul consensus recommends blastocyst assessment to be performed at 116 ± 1 hours post-insemination.^[5]

The importance of this parameter becomes evident when considering scenarios such as comparing a grade 4AA blastocyst on day 5 with a similar grade 4AA blastocyst on day 6. Associating these with a similar score would be inaccurate. For optimal training of AI models in blastocyst assessment, the ideal dataset should use images captured within a consistent time frame, ensuring greater accuracy and comparability of the results.

Ideal Endpoint or Ground Truth for Datasets Used for Blastocyst Assessment

The choice of endpoint in blastocyst assessment studies depends heavily on the primary aim of assessment. When the goal is to rank and prioritise blastocysts for transfer, an ideal endpoint would be the detection of cardiac activity, as it provides a more immediate indicator of viability compared to live birth, which is influenced by numerous confounding factors.^[54]

In studies aimed at predicting euploidy, the choice of endpoint is complex. Live birth serves as a comprehensive endpoint, encapsulating the ultimate goal of achieving a successful pregnancy. However, for cases resulting in abortion, assessing the genetic status of the conceptus using CVS can offer valuable insights. By doing so, we can gain a more accurate understanding of the predictive capabilities of the model.

Importance of Addressing Confounders in AI Model Training

Maternal factors such as very advanced maternal age (age > 40 years) and high BMI, and uterine factors such as adenomyosis, endometriosis, and the presence of uterine structural anomalies significantly affect the clinical pregnancy and live birth outcomes.^[55–57] It’s crucial to acknowledge that neglecting these confounders during dataset preparation may introduce biases that affect the robustness of AI model learning.

Training an AI model without accounting for these confounders can negatively affect its performance. For example, if an AI model learns from cases where a blastocyst with favourable morphology does not result in pregnancy due to underlying uterine abnormalities, it may lead to incorrect generalisations. This failure to consider the broader clinical context can impair the AI model’s understanding of the true relationship between embryo quality and outcomes. Such misleading data are labelled as ‘noisy data’ and have the ability to impact model performance.^[58]

Excluding confounders is vital not only for ensuring that the dataset represents clinical reality accurately but also for optimising the learning process of AI models, allowing them to capture the various factors associated only with blastocyst morphology that can influence reproductive outcomes.

Importance of Dataset Split or Class Distribution

Highlighting the significance of mirroring real-world scenarios in dataset class distribution, it is crucial to note that, typically, close to 40% of embryos are classified as good usable blastocysts, whereas close to 60% are categorised as poor blastocysts.^[59] A recent study suggested that altering the class distribution from 30:70 to 50:50 does not significantly affect the AI model performance.^[37] However, another study cautioned against oversampling the negative class (poor embryos), as it introduces a notable negative bias in the AI model.^[15] Thus, advocating for a closer-to-reality class distribution of approximately 40% good and 60% poor seems logical for robust AI training.

Deciding on the correct dataset split for training AI models in blastocyst assessment is a critical consideration. Here is a concise overview of the pros and cons associated with a 50%–50% split and a more practical 40%–60% split.

Importance of Embryo Transfer Strategy in Studies

In studies aiming to associate a predictive score with embryos or assess embryo viability during blastocyst assessment, the choice of embryo transfer strategy used during dataset collection has the potential to impact AI model performance. If the AI model is trained solely on fresh transfers, various confounders, like uterine receptivity and endometrial thickness at the time of transfer, may affect the ground truth. Failure to account for these confounders can introduce bias during model training. Therefore, advocating for an ideal strategy, it is suggested that using only FETs for dataset collection, avoiding a mix of fresh and FETs, could help mitigate confounding factors and enhance model robustness.

Importance of Using Clinical Metadata Along with Images

In the context of metadata utilised by models, it is essential to acknowledge that the implantation potential of an embryo is influenced by various factors, with maternal age being one of the most significant contributors.^[61] Models aimed at predicting implantation or assigning a viability score can greatly benefit from incorporating patient age and other relevant clinical parameters into their analysis.

Proposed Characteristics of a Gold Standard Dataset

A gold standard dataset for AI-based blastocyst assessment should be designed to support robust training, validation, and testing of clinically applicable models. Based on insights from existing AI tools, such a dataset should ideally include between 50,000 and 100,000 high-quality images of expanded blastocysts, sourced from multiple clinics to ensure diversity and generalisability. A minimum threshold of 20,000 well-curated images may serve as a baseline when image quality, patient selection and imaging consistency are tightly maintained. These images should be divided into 70% for training, 15% for validation and 15% for testing the AI model.^[62]

The dataset should consist exclusively of expanded blastocysts from autologous IVF cycles, with one embryo per image and no other embryos, microtools, or visual distractions in the frame. Each image must clearly display the entire blastocyst in sharp focus, with identifiable ICM, TE and zona pellucida. Image capture should be performed at a consistent time point between 110 and 116 hpi using a 20x objective lens on an inverted microscope, resulting in a total magnification of approximately 200 times. A minimum image resolution of 500 by 500 pixels should be maintained. The dataset should include patients aged between 23 and 38 years with normal BMI and exclude cases with uterine abnormalities such as endometriosis, adenomyosis, congenital anomalies, or recurrent implantation failure. Only frozen cycles should be included to eliminate variability due to endometrial receptivity. The ground truth for each image should be defined as the presence or absence of foetal cardiac activity between 6 and 8 weeks of gestation.

All images should be annotated with clinical outcome labels for classification purposes. In addition, a smaller subset of 500–2000 images should be structurally annotated using standard annotation tools to clearly mark the ICM, TE and zona pellucida. These detailed annotations are essential for training segmentation models or for developing region-focused explainable AI tools. A class distribution of approximately 40% positive and 60% negative outcomes should be maintained to reflect real world clinical conditions.