Effect of Data Augmentation and Lung Mask Segmentation for Automated Chest Radiograph Interpretation of Some Lung Diseases

Abstract

The results of chest X-ray (CXR) analysis of 2D images to get the statistically reliable predictions of some lung diseases by computer-aided diagnosis (CADx) based on the convolutional neural network (CNN) are presented for the largest open CXR dataset with radiologist-labeled reference standard evaluation sets (CheXpert). The results demonstrate the lower validation loss and higher area under curve (AUC) values for the receiver operating characteristic curve (ROC) for the models with lung mask segmentation (for 4 from 14 lung diseases) and data augmentation (for 10 from 14 lung diseases) for small image sizes (\(320\times 320\) pixels) and standard CNN (like DenseNet121) even. Moreover, the additional training leads to the lower validation loss and higher AUC values for the model with data augmentation. The further progress of CADx is assumed to be obtained for the big datasets with the bigger original image sizes by longer training with the tuned data augmentation.