SAMamba: A Segment Anything Model-Augmented Framework Integrating Mamba for Comprehensive Medical Image Segmentation

In the field of medical image segmentation, developing a high-performance, generalizable model is crucial for improving the accuracy of various medical image analysis tasks. We present SAMamba, a novel deep learning framework that integrates state space models (SSMs) for general-purpose segmentation. SAMamba employs a U-shaped architecture based on SSMs, utilizing Visual State Space (VSS) blocks as its core modules. This design effectively captures rich contextual information while an encoder-decoder structure facilitates feature extraction and multi-scale resolution integration. We evaluated SAMamba's performance across three distinct medical image datasets: the Breast Ultrasound Dataset (BUSI), the Thyroid Ultrasound Dataset (DDTI), and the ISIC 2018 Skin Lesion Dataset. Experimental results demonstrate that SAMamba delivers outstanding segmentation performance on all three datasets, particularly excelling at delineating complex lesion boundaries. Compared to traditional convolutional neural networks (CNNs) and Transformer-based models, SAMamba maintains high accuracy while reducing computational complexity.