Momentum-Guided Architecture Optimization for Trophectoderm Segmentation in Day-5 Human Blastocyst Images

In assisted reproduction, accurate segmentation of human blastocyst (HB) components is essential for embryo assessment and selection. Deep learning-based segmentation methods, including U-Net variants and DeepLab architectures, have demonstrated strong performance. While these algorithms have shown strong performance, their architectures are typically designed manually and require extensive hyperparameter tuning, which is prone to overfitting on small medical datasets. To address the challenges, this study proposes Adam-guided U-Net for blastocyst segmentation, an architecture search framework that replaces conventional fixed-shape architectures/parameter tuning with Adam-inspired adaptive first- and second-moment updates. Adam-Evolved U-Net combines Adam-inspired adaptive moment updates to explore the search space for a constrained yet flexible U-shaped backbone. The search space includes convolutional depth, filter configuration, and dropout rate. The search process begins with a randomly initialised U-Net configuration and iteratively evolves based on segmentation fitness. Unlike conventional mutation-based evolutionary approaches, Adam-Evolved U-Net uses moment-guided updates from the Adam optimizer for search stability and directional exploration in the search space. The optimal architecture identified is trained to segment Trophectoderm in day-5 HB microscopic images. Experimental results show that Adam-Evolved U-Net achieves better segmentation performance. Dice score of 89.69 and a Jaccard index of 81.75 outperform several existing segmentation approaches under low computational complexity. Additional experiments on inner cell mass and retinal vessel segmentation justify the adaptability and generalisation capability of the proposed framework for medical image segmentation.