An Efficient Polyp Detection Framework with Suspicious Targets Assisted Training

Automatic polyp detection during colonoscopy is beneficial for reducing the risk of colorectal cancer. However, due to the various shapes and sizes of polyps and the complex structures in the intestinal cavity, some normal tissues may display features similar to actual polyps. As a result, traditional object detection models are easily confused by such suspected target regions and lead to false-positive detection. In this work, we propose a multi-branch spatial attention mechanism based on the one-stage object detection framework, YOLOv4. Our model is further jointly optimized with a top likelihood and similarity to reduce false positives caused by suspected target regions. A similarity loss is further added to identify the suspected targets from real ones. We then introduce a Cross Stage Partial Connection mechanism to reduce the parameters. Our model is evaluated on the private colonic polyp dataset and the public MICCAI 2015 grand challenge dataset including the CVC-Clinic 2015 and Etis-Larib, both of the results show our model improves performance by a large margin and with less computational cost.