HMSA-SegNet: A Parameter-Efficient Architecture with Hierarchical Multi-Scale Attention for Medical Image Segmentation

Le résumé fourni par la source

Medical image segmentation plays a critical role in diagnostic and therapeutic applications, requiring high accuracy while maintaining computational efficiency. Traditional CNNbased methods such as U-Net effectively captured spatial features but often struggled with long-range dependencies. Transformerbased models improved global context modeling but introduced substantial computational overhead. To address these challenges, this study proposed HMSA-SegNet, a parameter-efficient with real-time inference segmentation network integrating Adaptive Multi-Path Dilated Convolution (AMDC), Transformer Interaction (TI), and Deformable Coordinate Attention (DCA). AMDC dynamically adjusted dilation scales through learnable weights, capturing multi-scale contextual information more effectively than conventional atrous convolutions. TI employed a single multi-head self-attention mechanism to enhance cross-scale feature interaction without excessive computational cost. DCA refined spatial and channel attention, improving boundary delineation by leveraging deformable convolutions. The proposed model was evaluated on CHAOS and ISIC-2018 datasets, demonstrating advanced segmentation performance while maintaining significantly fewer parameters. Additionally, experiments on ISIC-2018 highlighted its scalability to different imaging modalities, effectively handling complex lesion boundaries. The modular design of HMSA allowed plug-and-play integration into other segmentation architectures, making it a practical solution for real-time medical imaging applications. These findings suggested that HMSA-SegNet provided an optimal balance between accuracy, efficiency, and computational cost.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.