GM-UNet: Graph Mamba UNet for Medical Image Segmentation

Medical image segmentation stands as a pivotal element in the domain of computer-aided diagnosis, facilitating the accurate delineation of anatomical structures and pathological regions. Yet, the intricacy of medical images, coupled with the diverse appearance and shape of various anatomical and pathological entities, poses notable challenges. This study introduces the GM-UNet architecture, a novel framework that integrates Graph Cross Attention (GCA) and Parallel Channel Spatial Attention (PCSA) modules, aimed at enhancing the efficacy of medical image segmentation. The innovative GM-UNet architecture employs graph neural networks to capture global dependencies, fostering a more holistic comprehension of the image. Through the incorporation of GCA, our model adeptly addresses long-range dependencies among pixels, significantly boosting segmentation precision. Additionally, the PCSA modules are designed to concentrate attention mechanisms on pertinent features across spatial and channel dimensions, thereby augmenting segmentation outcomes. The performance of the GM-UNet architecture was rigorously evaluated across two distinct tasks: skin lesion segmentation leveraging the ISIC2017 and ISIC2018 datasets, and organ segmentation utilizing the Synapse dataset. The findings affirm that the GM-UNet architecture outperforms existing state-of-the-art models concerning segmentation accuracy, marking a notable advancement in the field of medical image analysis. The integration of GCA and PCSA modules within the GM-UNet architecture underscores the potential of our model to significantly contribute to precise and reliable medical diagnoses.