Real-time colorectal polyp detection via YOLOv10

Colonoscopy is a critical means for early colorectal polyp detection, which can significantly reduce colorectal cancer mortality. However, it relies on clinicians' expertise and is prone to errors due to subjective factors such as fatigue. To solve this problem, we propose a polyp detection method based on YOLOv10. Firstly, a small dataset of polyps is manually annotated to construct a pre-trained model. Then, a semi-automated iterative labeling strategy is applied to rapidly generate a larger labeled dataset. Next, the proposed polyp detection model is trained by the large dataset. Finally, the trained model is evaluated on a test set to identify polyp locations and provide labels and confidence scores via bounding boxes. The experimental results indicate that our model achieves a good balance between detection accuracy and speed, and can effectively detect colorectal polyps in real-time.