Classification of CXR Chest Diseases by Ensembling Deep Learning Models

Abstract

Chest related diseases are the most frequent health issues worldwide. They are mainly diagnosed by radiologists using a visual inspection of a chest X-ray (CXR) radiography. This task is challenging and error-prone because of the similarity between signs of the diseases, which occur as opacities around the infected organs. An early diagnosis of chest abnormalities is an essential step in the treatment process. Computer-aided detection (CAD) systems can be useful as decision support tools for radiologists. In recent years, chest disease detection using deep learning (DL) and CXR images has become one of the most promising topics in healthcare. In this work, we propose an approach using an ensemble of deep convolutional neural networks (DCNN) for a two-steps classification of CXR images. In the first step, a CXR image is taken as input of our model. This latter determines whether it is a heart or a lung abnormality. Then, it performs a binary classification for diseases of the same organ as a second step. A dataset of 26,316 CXR images was consolidated by merging images from two open-access datasets (VinDr-CXR and CheXpert). Our approach showed impressive results, obtaining an AUC of 0.9489 for infected organ classification (multi-class classification) and a mean AUC of 0.9957 for specific diseases' classification (binary classification) surpassing state-of-the-art DCNN models.