PMSDN: Progressive Multi-Scale Dual Network for Brain Vessel Segmentation

초록

We present PMSDN (Progressive Multi-Scale Dual Network), a dual-encoder architecture that combines a 2.5D U-Net and a Vision Transformer (ViT) for brain vessel segmentation. To overcome the limitations of conventional 2D and 3D models - namely, limited global context and high computational cost - PMSDN integrates ViT with maximum intensity projection (MIP) inputs to capture long-range dependencies, while enhancing the U-Net with lightweight attention modules for efficient local feature extraction. A unidirectional cross-attention mechanism fuses local and global features. PMSDN achieves a Dice Similarity Coefficient (DSC) of 0.904 on internal validation and demonstrates competitive performance on external datasets (ICBM: 0.764, LocH1: 0.826), surpassing or matching recent state-of-the-art models without retraining. These results highlight its potential for multi-institutional deployment. © 2025 IEEE.

키워드