Lung Disease Classification using Deep Learning and ROI-based Chest X-Ray Images

Abstract

[eng] —Automatic pathology detection in chest X-ray images is of paramount importance due to its potential to significantly enhance diagnostic accuracy and efficiency in clinical settings. With the increasing prevalence of lung diseases worldwide, there is a critical need for reliable and efficient methods to aid radiologists in early and accurate diagnosis. This study presents a contribution on automatic lung disease classification using deep learning techniques applied to chest X-ray images, addressing some of the most prevalent lung pathologies. The main focus is on the use of convolutional neural networks to identify and classify different lung diseases. The research compares two approaches: a direct multiclass classification model and a two-stage classification model. The two-stage approach, which first detects abnormalities and then classifies them into specific diseases, offered superior performance in reducing false positive rates for normal cases. However, both methods showed comparable metrics without statistically significant differences. Segmentation techniques were employed in the study to isolate lung regions, allowing the methods to focus solely on the regions of interest. Contrary to expectations, lung segmentation did not improve classification performance and instead led to worse results. This suggests that models trained on whole images may rely on database-specific rather than disease-related features. Data augmentation techniques were used to increase model robustness, but their impact on performance was minimal. The study also explored explainability methods such as GradCAM and Score-CAM to localize disease-specific regions within the radiographs, but these methods did not reliably identify the precise locations of certain pathologies, such as pulmonary fibrosis. In addition, experiments with bone suppression techniques, which aim to remove bone structures from X-rays, did not provide robust improvements in classification accuracy due to the degradation of image quality caused by resizing