Contrastive Feature Decoupling for Weakly-Supervised Disease Detection

Jhih-Ciang Wu, Ding-Jie Chen, Chiou-Shann Fuh

Abstract: Machine learning based Computer-Aided Diagnosis (CAD) aims to assist clinicians in the pathological diagnosis process. While dealing with video pathological diagnosis such as colonoscopy polyp detection, the recent SOTA method employs Weakly-supervised Video Anomaly Detection (WVAD) in the Multiple Instance Learning (MIL) scenarios to concern the temporal correlation within data and to formulate the concept of the interest disease simultaneously. Such a MIL-based WVAD method leverages video-level annotations to detect frame-level diseases and shows promising results. This paper casts the video pathological diagnosis as a MIL-based WVAD task and introduces Contrastive Feature Decoupling (CFD) network to decouple normal and abnormal feature ingredients per snippet. With such decoupled features, we are able to highlight the abnormal feature ingredients for accurately reasoning the disease score per snippet. The core components within our CFD model are the memory bank and contrastive loss. The former is used to learn atoms for representing normal features, and the latter is used to encourage our model to gain robust disease detection. We demonstrate that our CFD network is achieving new SOTA performance on the existing Polyp dataset and the introduced PANDA-MIL dataset. Our dataset are available at https://github.com/Jhih-Ciang/PANDA-MIL .