Automated Polyp Detection System in Colonoscopy using Object Detection Algorithm based on Deep Learning

In Korea, colon cancer is increasing due to westernized eating habits. Colonoscopy is being used to reduce deaths from colon cancer and studies of CADx(Computer-aided Diagnosis) are being developed to improve accuracy. Due to the nature of medical data, it was difficult to collect a lot of data, so data was increased 25 times using AutoAugment's CIFAR-10 policy, and YOLOv4(You Only Look Once), a real-time object detection algorithm, was used to detect lesions. A new object detection algorithm, YOLOv4, use new eight features such as Weighted-Residual-Connections, Cross-Stage-Partial-connections, Cross mini-Batch Normalization and Self-Adversarial-Training. The performance of augmented data had a maximum mAP of 27.44 higher than the original data. The average IoU(Intersection over Union) was 11.44 higher than the original data. When the IoU value is 0.5, the F1-scores of the original data and the augmented data are 0.9 and 0.97 respectively. © 2021 Korean Institute of Electrical Engineers. All rights reserved.