Improve Performance of Deep Learning Polyp Detection System in Clinical Colonoscopy

Background: Colonoscopy is the gold standard of colon cancer screening though some polyps are still missed, with polyps miss-rate (PMR) are estimated to be as high as 22%. Computer aided polyp detection may help colonoscopists reduce their polyps miss-rates.Aim: This study focuses on how to improve the performance of AI system for detecting polyps by adding false positive noise picture for learning.Methods: We focus our efforts of contracting the performance difference of test results in images dataset and video dataset, after adding noise images and before. Train the YOLOv3 model 1 using noise images. Train the YOLOv3 model 2&3 and RetinaNet model using noise images with different learning mode.Result: In this work, we proposed a high sensitivity training approach with low false-positive that can perform accurate detection with accuracy that reaches a false-positive of less than 10%. According to the image test result of two open datasets and Renmin hospital dataset, the YOLOv3-MODEL 1 model is very sensitive to polyps, but false-positive in non-polyps videos also reached 56.64%, it means excessive false-positive prompts will mislead endoscopists, thereby increase unnecessary inspection time. Unlike MODEL 1, we add negative noise samples in the training set when we trained YOLOv3-MODEL2, YOLOv3-MODEL3 and RetinaNet-MODEL. According to the video test set result, it can be seen that false positives are greatly reduced, meanwhile the decrease in sensitivity is also within the acceptable range. Among them YOLOv3-MODEL3 with mosaic data enhancement and RetinaNet-MODEL with non-maximum suppression both maintain more than 90% sensitivity, and false-positive is less than 10%, which means that the model can achieve the desire identification performance for practical application. This novel training approach allow improving detection rate and reducing misleading can be achieved at the same time.