GH-UNet: Group-wise Hybrid Convolution-VIT for Robust Medical Image Segmentation

Medical image segmentation is essential for accurate diagnosis and effective treatment. Although U-Net-based architectures have shown outstanding performance in this domain, they are limited in effectively capturing long-range contextual relationships within complex anatomical structures. Consequently, this study presents GH-UNet, a G roup-wise H ybrid Convolution-VIT model under the U-Net framework for robust medical image segmentation. GH-UNet is designed to efficiently capture local details and long-range dependencies while accurately delineating target boundaries in complex medical scenarios. First, we designed a hybrid convolution-vision Transformer (VIT) encoder to effectively capture both local and long-range dependencies. Second, a Group-wise Dynamic Gating (GDG) component was developed to dynamically adjust feature weights, thereby enhancing the effectiveness of feature expression. Third, we implemented a cascade mechanism in the decoder to integrate information across various scales. Notably, both the hybrid convolution-VIT encoder and the GDG component are modular, allowing for broad application across CNN or VIT architectures. Experimental results demonstrate that GH-UNet performs exceptionally well across five public and one private 2D/3D medical image segmentation datasets. For example, on the ISIC2016 dataset, GH-UNet outperformed the H2Former model, improving the DICE and IOU scores by 1.37% and 1.94% respectively, while requiring only 38% of the parameters and 49.61% of the FLOPs compared to H2Former. The source code is freely accessible via: https://github.com/xiachashuanghua/GH-UNet.