GAN-based Spatial Transformation Adversarial Method for Disease Classification on CXR Photographs by Smartphones

Deep learning has been successfully applied on Chest X-ray (CXR) images for disease classification. To support remote medical services (e.g., online diagnosis services), such systems can be deployed on smartphones by patients or doctors to take CXR photographs using the cameras on smartphones. However, photograph introduces visual artifacts such as blur, noises, light reflection, perspective transformation, moiré pattern, etc. plus unwanted background. Therefore, the classification accuracy of well-trained CNN models performed on the CXR photographs experiences drop significantly. Such challenge has not been solved properly in the literature. In this paper, we have compared various traditional image preprocessing methods on CXR photographs, including spatial transformation, background hiding, and various filtering methods. The combination of these methods can almost eliminate the negative impact of visual artifacts on the evaluation of 3 different single CNN models (Xception, DenseNet-121, Inception-v3), only 0.0018 AUC drop observed. However, such methods need user manually process the CXR photographs, which is inconvenient. Therefore, we have proposed a novel Generative Adversarial Network-based spatial transformation adversarial method (GAN-STAM) which can automatically transform the CXR region to the center and enlarge the CXR region in each CXR photograph, the classification accuracy has been significantly improved on CXR photographs from 0.8009 to 0.8653.