A Review on Deep Learning for Pulmonary Disease Classification Through CheXpert Dataset’s Chest X-Ray Images

Abstract

Chest radiography stands out as a common imaging modality for the screening and monitoring of pulmonary disease, but its interpretation is hindered by over-lapping presentations, image variability and inter-observer disagreement. Deep learning offers a path towards consistent, scalable computer-aided diagnosis, and public release of CheXpert dataset has catalysed a substantial body of method-ological work. This literature survey presents a systematic, PRISMA-guided review of deeplearning approaches developed and evaluated on CheXpert dataset for the multi-label classification of thoracic pathologies. From an initial pool of 197 records identified across IEEE Xplore, PubMed, Scopus, Web of Science, the ACM Digital Library and arXiv, 26 primary studies met the inclusion criteria and are analysed in depth. We organise the literature into a taxonomy cov-ering convolutional, transformer, self-supervised, vision–language, graph-based and multimodal architectures, and we compare reported performance on the five official competition pathologies (Atelectasis, Cardiomegaly, Consolidation, Edema and Pleural Effusion). We further examine how each study handles CheX-pert’s uncertainty labels, model calibration, robustness to image perturbations and explainability. The synthesis identifies persistent open problems—noisy and uncertain labels, demographic bias, weak external generalisability, the absence of standardised evaluation protocols, and limited clinical validation of reported gains—and outlines concrete future directions including foundation-model adap-tation, federated training, fairness auditing, calibration-aware learning and prospective clinical evaluation.