kaihealth | research

Back to list

Enhancing explainability of AI models in pregnancy prediction for IVF through edge-aware visual feature extraction from embryo images

Jun 23, 2025

Authors

H M Kim, Y Jo, M K Chung, J Kim, Y J Lee, S E Moon, H J Yang, J M Yim, H Lee

Conferences

ESHRE

ABSTRACT

Study question

Can Active Boundary Loss (ABL) improve AI models' ability to accurately identify ICM, TE, and ZONA boundaries, enhancing explainability and consistency in embryo analysis?

Summary answer

ABL enhances AI models' performance in recognizing ICM boundaries, improving accuracy, boundary detection, and transparency, fostering consistent embryo assessments and better clinical decision-making.

What is known already

A critical aspect of embryo assessment, whether performed by embryologists or AI, lies in evaluating the inner cell mass (ICM) and trophectoderm (TE), as these structures are most indicative of IVF success. However, the grading process is inherently subjective, largely due to the ambiguous boundaries of ICM and TE. This inconsistency highlights the need for AI models capable of accurately identifying ICM and TE boundaries, as well as the zona pellucida (ZONA), to enhance the reliability and explainability of embryo analysis. By improving boundary recognition, such models could support more precise and consistent decision-making in IVF.

Study design, size, duration

Between June 2015 and January 2022, 1,198 embryo images were collected from four IVF clinics in South Korea. Five embryologists manually annotated the objects of interest, including ICM, TE, and ZONA, at the pixel level. We trained AI models, including fully convolutional networks (FCN), DeepLab V3, and SegFormer B5 , for visual feature extraction using these pixel-level annotations. These images were split into training (60%), validation (20%), and testing (20%) sets.

Participants/materials, setting, methods

This study assessed the ABL function for improving boundary-aware visual feature extraction of objects of interest, including ICM, TE, and ZONA. ABL was used to enhance visual feature extraction performance. The AI models and loss functions were implemented using the PyTorch deep learning framework. Model performance was evaluated based on the mean intersection over union (MIoU) for ICM, TE, and ZONA, as well as boundary IoU to measure boundary detection precision.

Main results and the role of chance

Accurate recognition of ICM and TE boundaries is critical for reliable embryo assessment in IVF, as these structures significantly impact clinical outcomes. With ABL, AI models achieved notable improvements in visual feature extraction. The MIoU performance (%) for objects of interest (ICM, TE, and ZONA) using ABL was 70.75, 70.45, and 74.21 for FCN, DeepLab V3, and SegFormer B5, respectively, with SegFormer B5 achieving the highest performance. When trained with ABL, the IoUs for ICM, TE, and ZONA were 71.90, 69.22, and 81.53, compared to 69.19, 69.50, and 81.10 without ABL. Boundary IoU also improved from 55.66 (without ABL) to 56.31 (with ABL). These findings highlight the potential of ABL in enhancing visual feature extraction for complex and ambiguous structures like ICM and their boundaries. The improved visual feature extraction can enhance explainability in several ways. First, it enables embryologists to assess regions with ambiguous boundaries like ICM more consistently. Second, highlighting key regions, including ICM, TE, and ZONA, facilitate the communication with patients. Finally, the extracted ICM, TE and ZONA can also be used as input data for pregnancy prediction and embryo grading, enhancing the reliability of AI models in clinical IVF applications.

Limitations, reasons for caution

While ABL improved ICM recognition, TE and ZONA performance remained similar, likely because of variations in microscope magnification and conditions across IVF labs. To address this, we plan to adopt preprocessing techniques and design models robust to diverse imaging conditions, focusing on robust feature extraction across diverse conditions.

Wider implications of the findings

Our study suggests that visual features of embryos can be more accurately extracted when their boundaries are explicitly identified. We expect that extracting visual features and incorporating them into a pregnancy prediction model will enhance the interpretability of AI models, often regarded as black-box systems.