ARBS-YOLOv8 for Annotation Box Detection in Colonoscopy CADe Systems

Colonoscopy is the gold standard for colorectal lesion detection, with its accuracy directly influencing patient prog-nosis. The accurate detection of Annotation Boxes in Computer-Aided Detection (CADe) systems is important for improving their performance in colonoscopy procedures. This research introduces a novel algorithm that combines active learning with a modified YOLOv8 algorithm, named ARBS-YOLOv8 (Active learning with RepVGG Block and SIoU loss function base on YOLOv8), for the detection of Annotation Boxes generated by colonoscopy CADe systems. We compared standard YOLOv8, YOLOv6, YOLOX algorithms and their active learning variants using colonoscopy video data from 31 patients. The active learning method involves analyzing false positives generated by the initially trained model on an active learning dataset. These false detection samples are then reintroduced into the training set as negative examples. This active learning method led to substantial performance improvements across all algorithms, with particularly notable increase in precision. Notably, the application of active learning to YOLOv6 resulted in a significant increase in Precision by 23.28%, reaching 95.79%. Furthermore, when active learning was integrated with YOLOv8, it achieved an impressive F1 Score of 99.70% on the test set. Compared to the standard YOLOv8, ARBS-YOLOv8 demonstrated 0.88% increase in F1 Score while maintaining high Sensitivity. This improvement suggests that ARBS-YOLOv8 is more effective at extracting features from the data and increase the F1 score in detecting Annotation Boxes.