Segmentation-based Knowledge Extraction from Chest X-ray Images

Abstract

Computer-aided detection applications have been extensively used to assist physicians in clinical diagnoses. Extracted information from X-ray, positron emission tomography, and magnetic resonance images enables radiologists and other physicians to identify pathologies, correlate findings with the symptoms, and determine the treatment steps. In this study, we proposed an automatic knowledge extraction methodology from chest X-ray images. The extracted knowledge is obtained from the segmented sections of the images that include pathological findings. We evaluated these segmented images with a) classical machine learning and b) pretrained convolutional neural network (CNN) models. Evaluations were based on areas under the receiver operating characteristic (AUROC) with segmented images using the pretrained CNN and the traditional method models, and they produced the average AUROC scores of 0.96 and 0.52, respectively. Traditional methods yielded lower AUROC scores compared with pretrained CNN methods. However, traditional methods may still be considered as appropriate solutions for disease diagnoses primarily based on their advantages regarding running time and flexibility.