Abstract
Colon cancer is a prevalent malignancy, substantially it prevented most effectively from killing patients through early endoscopic detection. With the rapid development of artificial intelligence technology, the early diagnosis rate of colonic polyps achieves greater clinical efficacy for colon cancer by applying target detection algorithms to colonoscopy images. This paper presents two outcomes achieved through the application of the improved YOLOv5s algorithm with annotated microscopy images of clinical cases and publicly available polyp image data: (1) enhancement of the C3(Cross Stage Partial Networks) module with multiple layers to C3SE(Cross Stage Partial Networks with Squeeze-and-Excitation) via the attention mechanism SE (squeeze-and-excitation) and (2) fusion of higher-level features utilizing BiFPN (the weighted bi-directional feature pyramid network). Experimental comparisons are performed based on a new image dataset of colonic polyps among more than 6 target detection algorithms to validate the better detection capability. The tests indicate that the YOLOv5s + BiFPN and YOLOv5s-1st-2nd-C3SE models exhibit enhancements of detection capability compared to the YOLOv5 algorithm according to the main indicators of the mAP, accuracy, and recall. The YOLOv5s + SEBiFPN model demonstrate a substantial improvement over the YOLOv5s algorithm, and establishing a benchmark technology for advancing computer-assisted diagnostic systems is feasible.