Expert Fusion Network for Automated Blastocyst Morphology and IVF Decision Support

Conventional morphology-based grading of blastocysts is subjective and inconsistent. Although artificial intelligence and deep learning methods have im proved performance, they continue to struggle with generalization across imaging domains, low-contrast structures, and multi-scale feature integration, which limits clinical applicability and may inflate performance estimates. We propose the Expert Fusion Network (EFN), an AI-driven framework designed to address these challenges in blastocyst assessment. EFN automates the classification and segmentation of the key embryonic structures: the trophectoderm (TE), blastocoel (BC), and inner cell mass (ICM), enabling precise morphological evaluation. A U-Net back bone delineates embryonic substructures with accurate boundary detection, while attention-guided feature learning enhances representation in ambiguous regions, particularly the low-contrast ICM. To capture structural heterogeneity, a mixture-of-experts (MoE) strategy is integrated, allowing specialized subnetworks to model TE, BC, and ICM individually. Experiments on 2,344 annotated images demonstrate substantial improvements over baseline mod els, achieving F1-scores of 0.96 for TE and BC, and 0.90 for ICM. Dice Similarity Coefficients (DSC) reach 0.92, 0.91, and 0.90 for TE, BC, and ICM, respectively. Cross-dataset generalization on an independent dataset yields F1-scores of 0.96, 0.97, and 0.91 for the same structures. An ablation study further validates the contribution of each architectural component. EFN demonstrates significant potential as a decision-support tool for in vitro fertilization (IVF), offering objective and reproducible assessments aligned with the Gardner grading system; however, prospective clinical validation remains necessary to establish its reliability and generalizability.