Deep learning based gastric lesion classification system using data augmentation

초록

Gastrointestinal symptoms and functional gastrointestinal disorders comprise a large proportion of primary care and gastroenterology practice. We propose a Computer-aided Diagnosis (CADx) system that analyzing the traditional gastroscope images and help the medical experts improve the accuracy of medical diagnosis. To improve the performance of the CADx system, a data augmentation has also been implemented to increase both the amount and the diversity of the training images. Augmentation method finds the enhancement parameters through RNN through large-scale verified three data, ImageNet, SHVN and CIFAR-10. In this study, we compared the performance of applying data augmentation method using four networks, Inception-V3, Resnet-101, Xception, and Inception-Resnet-V2. For Inception-V3, Resnet-101, Xception, and Inception-Resnet-V2 in normal and abnormal classification, the highest Az values were 0.87, 0.85, 0.88 and 0.82 respectively. The Xception networks and CIFAR-10 data is a promising CADx configuration for gastric lesion which had relatively simple structure and good classification performance. © Korean Institute of Electrical Engineers

키워드