GSAC-UFormer: Groupwise Self-Attention Convolutional Transformer-Based UNet for Medical Image Segmentation

Traditional transformers struggle to effectively capture local contextual information. Conversely, CNNs face challenges in modeling long-range dependencies. To address these limitations, this paper introduces GSAC-UFormer, an innovative Groupwise Self-Attention Convolutional Transformer-based UNet for medical image segmentation. The design of GSAC-UFormer focuses on efficiently integrating both local and global information, balancing the strengths of different processing techniques. At the core of GSAC-UFormer is the GSAC-Former block. This module combines groupwise convolution with a CNN-adaptive self-attention mechanism, enabling parallel integration of local and global contexts. This architecture allows the model to effectively capture intricate dependencies across various data dimensions while processing local features with high efficiency. The Guided Contextual Feature Attention (GCFA) mechanism further enhances feature selection. It emphasizes the most relevant contextual information, refining spatial and channel-wise relationships in the extracted features. This targeted approach mitigates noise and improves model accuracy. Additionally, the Multi-Depth Partitioned Depthwise Convolution Transformer (MDPDC-Former) serves as a bottleneck module. It optimizes feature mapping and enhances network learning efficiency by dynamically adjusting the receptive field. This enables the model to capture multi-scale semantic information more effectively. Experimental results highlight the superior performance of GSAC-UFormer compared to state-of-the-art methods. It achieves Dice coefficients of 91.6%, 94.61%, and 82.24% on the MICCAI 2017 (red lesion), PH2, and CVC-ClinicalDB datasets, respectively. These results underscore its effectiveness in advancing medical image segmentation.