CLN: a multi-task deep neural network for chest X-ray image localisation and classification

Chest X-ray (CXR) imaging is a widely used and cost-effective medical imaging technique for detecting various pathologies. However, accurate interpretation of CXR images is a challenging and time-consuming task that requires expert radiologists. Although deep learning methods have demonstrated high performance in CXR image classification, concerns over interpretability limit their clinical adoption. Localising pathologies on chest X-rays could improve interpretability and trust in these systems. In this work, we propose the Chest X-ray Localisation Network (CLN), a multi-task deep neural network designed to localise and classify pathologies in CXR images. Our proposed architecture was trained and evaluated on a subset of the ChestX-ray14 CXR data set, which included bounding box annotations of eight different pathologies from expert radiologists, achieving a maximum classification mean AUC score of 0.918 and a maximum localisation mean IoU accuracy of 0.855 for the eight examined pathologies (atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, and pneumothorax). Our approach outperformed state-of-the-art methods, demonstrating its potential as a reliable solution for computer-aided CXR image diagnosis, offering notable advantages over existing methods, including superior classification and localisation accuracy, reduced performance decay with increased IoU thresholds, and an overall simpler architecture.