MambaFormer: State Transformer Attention Network for Accurate Polyp Segmentation from Colonoscopy Images

Polyp detection plays an important role in preventing colorectal and gastric cancer. However, it is difficult to detect polyps in the colorectal and gastric cavities due to the following problems: 1) Lesion polyps are similar to adjacent normal tissue; 2) Some polyps are small and closely attached to the tissue. To address these challenges, a State Transformer Attention Network (MambaFormer) is proposed for polyp segmentation, where the Channel State Transfer Attention (CSTA) module is designed to distinguish lesion polyps from similar normal tissue by learning long-range dependencies between the lesion and normal regions, and the Space State Inference Attention (SSIA) module is developed to capture small polyps by focusing on the most relevant features of the objects. Experiments on benchmarks demonstrate the superiority of the proposed MambaFormer for polyp detection, achieving mean Intersection over Union (mIoU) scores of 89.3%, 94.7%, 73.6%, 90.2%, and 83.5% on the Kvasir, CVC-ClinicDB, CVC-ColonDB, EndoScene, and ETIS datasets, respectively.