BMS³: Bayesian Modeling Based SwinUNet Segmentation on Self-Distillation Architecture
· Paper · Medical Image Segmentation / Bayesian Modeling / SwinUNet / Self-Distillation
Published at ICONIP 2025.
Medical-image segmentation models often lose accuracy when scans come from a different device or acquisition setting. BMS³ addresses this problem by separating domain-invariant shape information from domain-specific appearance features with Bayesian modeling, then performing segmentation with a Swin Transformer-based U-Net.
An optional self-distillation objective improves learning on imbalanced datasets. Across multiple medical-imaging datasets, the method outperformed comparison systems including ResNet, TransUNet, and BayeSeg in segmentation accuracy and computational efficiency, while retaining stronger performance across different imaging domains.