BiF³-Net: A Full BiFormer Full-scale Fusion Network for Accurate Gastrointestinal Images Segmentation

Yunze Wang, Silin Chen, xi long, Yi.Tian, ye.huang, Tianyang Wang, Jingxin Liu

Keywords: Content-aware Sparse Attention, Full-scale Fusion, Over-Segmentation

Abstract: UNet-like segmentation models have been widely explored for the computer-aided segmentation and diagnosis of gastrointestinal (GI) tract diseases. However, the UNet architecture encounters two primary challenges: limited receptive fields due to conventional convolution operations, and a semantic gap arising from simplistic skip connections. In this paper, we introduce BiF³-Net, a novel model that integrates BiFormer blocks throughout the encoder and decoder, along with a full-scale BiFormer Fusion Bridge (BFB) module, aimed at addressing the aforementioned limitations. Meanwhile, we propose the Dense Inception Classifier (DIC) module to mitigate the over-segmentation problem in non-organ images. Extensive experiments demonstrate the effectiveness and adaptability of the proposed model.

Submission Number: 107