RS-FFGAN:Generative adversarial network based on real sample feature fusion for pediatric CXR image data enhancement

Abstract

With the increasing demand for computer-aided diagnosis technology in the medical field, convolutional neural network (CNN) has been widely used in recent years. The training of CNNs requires a large number of medical images. However, the pediatric CXR data are difficult to obtain and often suffer from imbalance problems. To augment the image data, the researchers augment the dataset with flipping, scaling, and increasing contrast. But this will lead to the problem of single samples, resulting in a large number of samples with no application value. In this paper, we propose a generative adversarial networks based on real sample feature fusion, called RS-FFGAN. It can synthesize realistic pediatric CXR and improve the quality of the generated composite image, solve the problem of unstable training, and more importantly, improve the prediction accuracy of the classification model under the condition of unbalanced data. We detect the quality and diversity of synthetic images by using IS and FID, the RS-FFGAN synthetic image has the lowest FID value, which is 32.73 for the Normal class and 57.55 for the Pneumonia class. The mean IS is the highest with 1.92 for the Normal class and 1.80 for the Pneumonia class. On small data sets and unbalanced data sets, using RS-FFGAN can improve the prediction accuracy of classification model by 5.88% and 15.62% respectively. The results show that using RS-FFGAN to synthesize images can balance the real data distribution and significantly improve the classification performance of CNN.