A Deep Learning Approach for Multi-Label Chest X-ray Diagnosis Using DenseNet-121

Abstract

This study addresses the increasing workload and diagnostic subjectivity in medical imaging by proposing a deep learning model for classifying chest X-rays into 14 distinct thoracic pathologies. Utilizing the DenseNet-121 architecture, A dataset of 112,120 frontal-view X-ray pictures is used to train the model. High Area Under the Curve (AUC) values are attained using the model for pathologies such as emphysema (0.93), Cardiomegaly (0.90) and pneumothorax (0.89), and performs comparably to State Of The Art(SOTA) models, including CheXNeXt. Our model demonstrates the potential to enhance medical image analysis by providing accurate, multi-pathology diagnosis from single X-ray images, thereby supporting radiologists in clinical decision-making. Future work will focus on improving performance for challenging pathologies like pneumonia (AUC 0.68) and consolidation (AUC 0.72). This model offers significant benefits for clinical workflows, perhaps lowering medical expenses and improving the results for patients.