Enhancing Cross-Domain Adaptability of Existing Computer-Aided Endoscopic Lesion Detection Using Plug-and-Play Tracker

Computer-aided detection (CADe) for endoscopy can help physicians to locate and identify lesions better, but there are still many false positives (FP) when processing cross-domain data. This paper proposes SE-SORT, a plug-and-play tracker, designed to be seamlessly integrated as a post-processing plugin into existing CADe systems to enhance the adaptability to cross-domain data. The proposed tracker adds trajectory initialization thresholds into the tracking association strategy, reducing the impact of high confidence FPs on the matching process. Experiments show that the modified tracker effectively reduces the impact without significantly affecting the processing speed. This allows the detector of CADe system to tolerate lower detection confidence thresholds, thus improving the overall accuracy on cross-domain data. Compared to existing SORT trackers, the proposed tracker exhibits better accuracy and higher efficiency in endoscopic lesion detection and tracking. This work will help to improve the generalization and expanding the clinical application scope of related works on endoscopic real-time CADe.