OCCMNet: Occlusion-Aware Class Characteristic Mining Network for multi-class artifacts detection in endoscopy.

Abstract

Multi-class endoscope artifacts detection is crucial for eliminating interference caused by artifacts during clinical examinations and reducing the rate of misdiagnosis and missed diagnoses by physicians. However, this task remains challenging such as data imbalance, similarity, and occlusion among artifacts. To overcome these challenges, we propose an Occlusion-Aware Class Characteristic Mining Network (OCCMNet) to detect eight classes of artifacts in endoscope simultaneously. The OCCMNet comprises the following: (1) A Dual-Branch Class Rebalancing Module (DCRM) rebalances the impact of various classes by fully exploiting the benefits of two complementary data distributions, sampling and detecting from the majority and minority classes respectively. (2) A Class Discrimination Enhancement Module (CDEM) effectively enhances the discrepancy of inter-class by enhance important information and introduce nuance information nonlinearly. (3) A Global Occlusion-Aware Module (GOAM) infers the obscured part of the artifacts by capturing the global information to initially identify the obscured artifacts and combining local details to sense the overall structure of the artifacts. Our OCCMNet has been validated on a public dataset (EndoCV2020). Compared to the latest methods in both medical and computer vision detection, our approach demonstrated 3.5-6.5% improvement in mAP50. The results proved the superiority of our OCCMNet in multi-class endoscopic artifact detection and demonstrated its great potential in reducing clinical interference.