Automated Classification and Localization of Thoracic Diseases from X- Ray Images Using Deep Learning
No Thumbnail Available
Date
2025-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
In developing countries, thorax diseases are the common cause of death. Chest X-ray (CXR) is the most common and effective procedure for the diagnosis of a wide variety of thorax diseases. However, an efficient chest x-ray interpretation requires experienced radiologists, which is particularly lacking in a developing country such as Ethiopia. Efficient computer-aided detection (CAD) of Chest X-ray that can assist radiologists to diagnose lung diseases and offer huge benefits to public healthcare. Currently ChestX-ray14 is the largest public dataset for CXR that includes 14 different pulmonary diseases and one normal class (such as atelectasis, cardiomegaly, consolidation, edema, effusions, emphysema, Fibrosis, Hernia, infiltrations, masses, nodules, no-finding, pneumonia, pneumothorax and Pleural Thickening); This large data set has opened the possibility of creating a better CAD system in radiology. In this study, the proposed Automated Thorax disease Classification and localization (ATDCL) has four main stages, the preprocessing stage, the segmentation stage, the classification stage and the localization stage. Image scaling, normalization, histogram equalization, and median filtering were employed in the preprocessing stage to improve the quality and remove noise from the input images. We have used U-net deep neural network for segmentation. The network composes of two paths: the down-sampling path and the up-sampling path, also known as encoder and decoder. The encoder part used to extract features from the image and then from these latent features, the decoder path will learn to reconstruct the high-resolution binary mask. Then the region of interest was taken out from the chest x-ray image to utilize as input for the next sub stages. In the thorax classification stage, we employed the inception model, a novel deep learning approach. We have achieved the Average accuracy value (Acc) of 0.8489 and we have used a Class Activation Map (CAM) method that can significantly improve the understanding of radiologists about the approximate location of the existed pathology. Our approach can provide radiologists a way to make quicker and more reliable decisions, and providing greater health care system.
Description
Keywords
Automated Thorax disease Classification and localization (ATDCL), Chest X-ray (CXR), Deep Learning (DL), Receiver Operating Characteristic (ROC), Area Under Curve (AUC), Class Activation Map (CAM), Computer-aided Detection (CAD), Convolutional Neural Network (CNN), deep neural network (DNN).