MFBGR: Multi-scale feature boundary graph reasoning network for polyp segmentation

At present, adding Transformer to CNN has promoted the rapid development of colorectal polyp image processing. However, from the perspective of multi-scale feature interaction and boundary coherence, there are mainly some limitations: (1) ignore the local and global correlation within the scale feature, which may cause the missed detection of tiny polyps, (2) lack of multi-scale features to explore the target region, which hinders the learning of multi-variant polyps, and (3) the semantic connection between the target area and the boundary is ignored, cause incoherent segmentation boundaries. In this regard, we design a multi-scale feature boundary graph inference network for polyp segmentation, namely MFBGR. First, the Transformer block captures local–global cues inside the multi-scale information learned by the CNN branches. Second, for the multi-scale global information generated by the Transformer block, we design a cross-scale feature fusion module (CSFM). CSFM performs scale-variation interaction and cascaded fusion to capture the correlation between features across scales and solve the scale-variation problem of segmented objects. Finally, the traditional boundary refinement or enhancement idea is generalized to the graph convolutional reasoning layer (BGRM). BGRM receives CNN's low-level feature information and CSFM's fusion features, or intermediate prediction results, and propagates cross-domain feature information between graph vertices, explores information between target regions and boundary regions, and achieves more accurate boundary segmentation. On the CVC-300, CVC-ClinicDB, CVC-ColonDB, Kvasir-SEG, ETIS datasets, MFBGR and mainstream polyp segmentation networks were compared and tested. MFBGR achieved good results, and Dice, IOU, BAcc, and Haudo were the best. The values reached 94.16%, 89.35% and 97.42%, 3.7442, and the segmentation accuracy of colorectal polyp images has been improved to a certain extent.