Switch-UMamba: Dynamic scanning vision Mamba UNet for medical image segmentation

Recently, State Space Models (SSMs), particularly the Mamba-based framework, have demonstrated exceptional performance in medical image segmentation. This is attributed to their capacity to capture long-range dependencies efficiently with linear computational complexity. Nonetheless, current Mamba-based models encounter challenges in preserving the spatial context of 2D visual features, which is a consequence of their reliance on static 1D selective scanning patterns. In this study, we present Switch-UMamba, an innovative hybrid UNet framework that integrates local feature extraction power of Convolutional Neural Networks (CNNs) with the abilities of SSMs for capturing the long-range dependency. Switch-UMamba capitalizes on the Switch Visual State Space (VSS) module to leverage the Mixture-of-Scans (MoS) approach, a new scanning mechanism that amalgamates diverse scanning policies by considering each scan head as an expert within the Mixture-of-Experts (MoE) framework. MoS employs a router to dynamically allocate appropriate scanning policies and corresponding scan heads for each sample. This sparse-activated dynamic scanning approach not only ensures a rich and comprehensive acquisition of spatial information but also curtails computational expenses. Our comprehensive experimental evaluation on several medical image segmentation benchmarks indicates that Switch-UMamba has achieved state-of-the-art performances without using any pretrained weights. It is also worth highlighting that our approach outperforms other Mamba-based models with fewer parameters.