Improvement of Colon Polyp Detection Performance by Modifying the Multi-scale Network Structure and Data Augmentation

This study proposed a computer-assisted diagnosis system that detects polyps during colonoscopy using a multiscale network structure. Medical data require institutional review board approval, and collecting sufficient data is challenging for several reasons. The amount of data may be small thereby resulting in overfitting. This study attempted to increase the amount of data available to solve this problem. Autoaugment and the policy applied to the CIFAR-10 dataset were used. This data augmentation can be learned immediately without review by a colonist because no changes in the shape of the polyp occur during colonoscopy with minimal movement in location. The object detection network used was YOLOv4, which is capable of multiscale learning. Multiscale learning is advantageous in detecting an object regardless of the size of the lesion because it can extract features of various sizes through one learning. In this study, the learning advantages of multiple scales were reinforced via the addition of scales to YOLOv4, while the learning accuracy was improved by changing the activation function. Therefore, the changed activation function can continuously extract features when updating the layer weight. When using all the methods presented, mAP exhibited the highest performance at 98.36.