HLSM-UNet:Hybrid Local Spatial Mamba UNet for Medical Image Segmentation
Xude Zhang, Xiaoping Wu, Changzhen Zhang, Yumin Tang, Dihong Luo, Houbing Tang · 2024
In the field of Medical Image Segmentation, deep learning models, especially Convolutional Neural Networks (CNN) and Vision Transformer (ViT) models, have been widely researched and applied in recent years. However, these two mainstream methods each face some challenges. CNN occupies a place in medical image processing with its powerful local feature extraction ability, but it is inadequate in dealing with global or remote feature dependencies, which to some extent limits its performance in complex medical image segmentation tasks. The ViT model, through its Self-Attention mechanism, can effectively capture long-range dependencies in images, but it also comes with high computational costs, especially the secondary computational complexity, which becomes a bottleneck that cannot be ignored when processing high-resolution medical images. In this paper, we propose a medical image segmentation method called HLSM-UNet based on a State Space Model(SSM). This model combines the advantages of UNet with the remote modeling capability of the Mamba model, aiming to achieve more efficient and accurate medical image segmentation. We conducted experiments on the publicly available ACDC MRI heart segmentation dataset and Prostate dataset, and the results showed that HLSM UNet outperforms existing CNN and Transformer models in multiple evaluation metrics. Especially in terms of remote modeling capability and computational efficiency, significant advantages have been demonstrated. This not only proves the effectiveness of the state space model in medical image segmentation, but also provides new ideas and methods for subsequent research.