An Improved Vision Mamba-Based Nested U-Net++ Architecture for Medical Image Segmentation

Medical image segmentation plays an important role in computer-aided diagnosis by helping clinicians to identify anatomical structures and pathological regions accurately. Encoder-decoder approaches like UNet and UNet++ have achieved best performances due to their accurate extraction capabilities of feature and multi-scale fusion information. But these approaches often got stuck because of their lack of efficiency to understand global contextual information, inaccurate boundary localization and limited robustness when they are associated with complex medical images with varying contrast and noise levels. These advancements motivated to propose an improved Nested Vision Mamba UNet++ (VMUNet++) architecture that integrates the latest selective state-space (S6) modules where nested encoder-decoder techniques applied. The effectiveness of the proposed model has been tested on several publicly available benchmark datasets like ISIC2018, Synapse, and SegPC-2021. Experimental outcomes have been evaluated by some standard popular segmentation metrics like Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Boundary F1 Score (BF Score), Average Surface Distance (ASD), and Hausdorff Distance (HD95) and superiority of the proposed technique has been noted.