LHP-VMambaNet: Lightweight Hybrid Parallel Vision Mamba Network for Efficient and Accurate Medical Image Segmentation

Accurate segmentation of medical images is essential for computer-aided diagnosis. Although both Transformer-based and CNN-based methods have demonstrated encouraging outcomes in medical imaging tasks, Transformers are limited by their quadratic computational cost, while CNNs are limited in their ability to simulate long-range dependencies. Establishing global contextual relationships is vital for successful segmentation. State Space Models (SSMs), emplified by Mamba, excel not only at modeling long-range interactions but also compensate for locally extracted features from convolutions and it maintains linear computational complexity. Moreover, multi-scale feature fusion and feature channel selection are key aspects of medical image segmentation that SSMs overlook. To this end, we propose a Lightweight Hybrid Parallel Vision Mamba Network (LHP-VMambaNet) for efficient and accurate medical image segmentation. Comprehensive experiments conducted on multiple diverse medical image datasets demonstrate that LHP-VMambaNet achieves competitive performance in medical image segmentation tasks. At the same time, the LHP-VMambaNet maintains high computational efficiency in terms of both parameter count and GFLOPs.