Classification of chest diseases from x-ray images on the chexpert dataset

Abstract

This work proposes a method to classify tuberculosis (TB) disease in a chest radiograph using convolutional neural network (CNN) algorithms. The main contribution of this work is to detect and classify TB disease in addition to the other five different diseases. This is achieved by using a transfer learning technique that utilizes a pre-trained CNN network to classify the TB disease. A comprehensive verification using TensorFlow is carried out to train and validate the proposed technique. This work aimed to use different pre-trained models on the CheXpert dataset and compare the area under the curve (AUC) between the CNN models. From the simulations, it was found that it is possible to classify the TB disease in addition to the other five diseases without having a degradation in the accuracy. The results confirm that transfer learning technique is superior to other methods, which exhibits less time for training and validating the datasets, and has good performance. This work achieved excellent performance in classifying three different diseases (atelec-tasis, edema, and tuberculosis) with AUC of 0.912, 0.945, and 0.954, respectively. Also, this work achieved second-best performance for classifying pleural effusion and consolidation diseases with AUC of 0.928 and 0.917, respectively. The method proposed in this work can be used for classification of diseases in chest radiograph as an early diagnosis tool in a clinical environment.