RMPNet: An Efficient Multi-Scale Colon Polyp Segmentation Network

Detection of intestinal polyps is one of the important means to prevent cancer and other pathological deteriorations, and early diagnosis is crucial for the treatment and prognosis of cancer. The polyp segmentation based on deep learning algorithms can identify and analyze the lesion characteristics in ultrasound images at the pixel level, providing doctors with references and basis. Therefore, it has important significance and value. Based on the encoder architecture of PVT (Pyramid Vision Transformer), this paper proposes an effective multi-scale feature recognition network, RMP-Net (RCA, MFE, PVTv2), which has achieved good results in various aspects. On the internal datasets CVC-ClinicDB and Kvasir, as well as the challenging external datasets CVC-ColonDB and ETIS-LaribPolypDB, Dice scores of 0.931, 0.924, 0.827, and 0.807 were achieved, respectively. The comprehensive performance is better than other networks compared in this paper.