Boundary-Aware Collaborative Mixture-of-Experts Architecture for Fetal Brain Anatomical Segmentation

Abstract

Accurate segmentation of fetal brain anatomical structures is crucial for prenatal disease screening and diagnosis. However, ultrasound images often suffer from noise, artifacts, and blurred boundaries, making manual delineation labor-intensive and prone to variability. Lightweight automated methods struggle with these challenges, while more complex algorithms are difficult to deploy on resource-constrained portable ultrasound devices and consumer-grade medical electronics. To address these issues, we propose the Boundary-aware Collaborative Mixture-of-Experts (BAC-MoE) framework. This approach dynamically adjusts the number of experts to optimize performance for resource-limited environments. It integrates frequency-domain filtering to enhance boundaries and textures, and spatial inverse attention to focus on small-scale, blurred structures. A collaborative learning-driven hybrid feature fusion strategy is designed to efficiently integrate complementary information from different expert branches, ensuring robustness while maintaining computational efficiency. Experiments on a fetal brain ultrasound dataset show that BAC-MoE significantly outperforms existing methods in segmenting tiny structures, achieving an optimal balance between accuracy and efficiency. These results demonstrate the potential of BAC-MoE for both clinical applications and deployment on consumer-grade ultrasound devices. The code is publicly available at https://github.com/201610006/BAC-MoE