| Title |
A Study on Data Augmentation Methods Optimized for Gastric Cancer Classification in Gastroscopy Images |
| Authors |
이정남(Jeong-nam Lee) ; 조현진(Hyun Chin Cho) ; 조현종(Hyun-chong Cho) |
| DOI |
https://doi.org/10.5370/KIEE.2021.70.12.2015 |
| Keywords |
CADx; Classification; Data augment; Deep learning; Gastric cancer |
| Abstract |
Gastric cancer is the most common cancer in Korea and an effective way to treat gastric cancer is early treatment. Gastroscopy is being performed for early detection of gastric cancer, and this paper proposes Computer-aided Diagnostics(CADx) system that can help gastroscopy. As a model for classifying gastric cancer, we use Xception, which reduces computation and improves performance.
Due to the nature of medical images, data augmentation was used to solve the lack of data and overfitting could occur. The data augmentation methods used were AutoAugment and Variational AutoEncoder(VAE). AutoAugment is a data augmentation method using color changes, shear, rotation, etc., and VAE learns to standardize the probability distribution of the data, helping to generate data similar to the original data to capture features. The data augmentation method with the best performance is the augmentation method through VAE, and the recall showed 94.84% of the performance. The data augmentation method optimized for the gastroscopy data set is the augmentation method through VAE. It can help the endoscopy specialist diagnose and increase the gastric cancer complete cure rate. |