An Accurate Polyp Segmentation Framework via Feature Secondary Fusion

Yanzhou Su, Qingsong Xie, Jin Ye, Junjun He, Jian Cheng

Abstract: Pre-segmentation of potential polyps can effectively improve the diagnostic efficiency of clinical specialists and reduce misdiagnosis and missed diagnosis. A common practice in polyp segmentation is to use Feature Pyramid Network (FPN) for segmentation. The vanilla FPN uses an extremely simple element-wise summation approach for feature fusion, resulting in less effective feature interactions. To solve the above issue, we provide a Feature Secondary Fusion Module (FSFM) in FPN to boost the performance of polyp segmentation. Specifically, in the first fusion, we form a dual-branch architecture to incorporate low-level spatial information into the high-level features and inject high-level semantics into low-level features, respectively, which can significantly reduce information loss. In the second fusion, we apply the gate mechanism to select the informative feature, making the feature fusion more effective. Moreover, rigorous evaluations are carried out on multiple polyp segmentation benchmarks. According to the experimental findings, integrating FSFM into a feature pyramid network surpasses other cutting-edge approaches.