BAFNet: Deep contour-aware features for colorectal polyps segmentation

Dibin Zhou, Ni Chen, Yueping Zhu, Xueyan Zhang, Innocent Nyalala, Jiayu Zhang, Junfeng Gao

Abstract: Feature channel selection has proven to be an effective strategy for establishing a top-down neural attention mechanism that refines and enhances semantic features for central representation. The quality of the semantic segmentation results for colorectal polyps is significantly influenced by both the encoder and decoder, with the encoder playing a pivotal inductive role in guiding the decoder’s final output. To improve the semantic accuracy of colorectal polyp segmentation in medical image-processing tasks, we introduce a boundary-aware feature fusion neural network (BAFNet). BAFNet incorporates two novel approaches. First, a gating branch was designed to construct an inhibition mechanism that selects high-quality contour features. This branch integrates a gating module with four convolutional-Batch Normalization (conv-BN) blocks. Second, to further construct prominent semantic features, we propose and incorporate the multi-scale feature aggregation module (MFAM) and dynamic global attention module (DGAM) as a robust feature constructor, parsing the semantics as a foreground representation. When tested on the CVC-ColonDB, ETIS, and EndoScene benchmark datasets, the proposed method achieved DICE scores of 0.808, 0.808, and 0.914; IoU scores of 0.722, 0.732, and 0.863; and F-measure scores of 0.782, 0.774, and 0.905, respectively, demonstrating a balanced and robust performance in colorectal polyp segmentation. Our neural network, with 5.98 million parameters and 3.59 GFLOPs, significantly outperforms ASCNet, MSNet, TransFuse, C2F-Net, and PraNet. These results demonstrate that the proposed algorithm is highly suitable for colorectal polyp segmentation applications and provides a robust solution for advancing medical image processing technology.

External IDs:dblp:journals/eswa/ZhouCZZNZG26