WCE‐DCGAN: A data augmentation method based on wireless capsule endoscopy images for gastrointestinal disease detection

Abstract

Abstract Wireless capsule endoscopy (WCE) is becoming more popular in clinical settings as a safe and painless gastrointestinal examination. Existing studies on automatic detection of lesions in WCE images have the problems of small dataset size and uneven distribution of numbers in terms of categories, which often leads to overfitting of the model and severely limits the performance improvement of the object detection network on WCE images. The traditional data enhancement methods such as flipping and local erasure have limitations and cannot achieve good generalization results. Therefore, a WCE‐DCGAN network was proposed in this paper to generate WCE images from existing WCE images. Using the images generated by this network and the original images as the input of the object detection network, there are different degrees of performance improvement on SSD, YOLOv5, and YOLOv4, and the average recognition accuracy of 97.25% can be achieved on SSD. Meanwhile, images generated by WCE‐DCGAN not only enlarge the size of the data set, but also have the characteristics of diversity, which makes the model have a good generalization effect.

Keywords

• data augmentation
• deep convolutional generative adversarial networks (DCGAN)
• generative adversarial network
• wireless capsule endoscopy (WCE) detection
• WCE‐DCGAN