A Deep Learning Approach to Automated Classification of Thoracic Diseases from Medical Imaging

Abstract

Thoracic diseases such as tuberculosis, pulmonary infiltration, and pleural effusion pose significant global health challenges due to their high prevalence and associated mortality. Early and accurate diagnosis is essential for effective management; however, traditional methods relying on manual chest radiograph interpretation are limited by expert availability, long analysis times, and inter-observer variability—especially in resource-constrained settings. This study presents a deep learning-based framework for the automated detection and classification of thoracic diseases from chest radiographs. The system uses a two-stage pipeline: a binary classification stage to detect tuberculosis, followed by a multi-class stage to distinguish between pulmonary infiltration, pleural effusion, and normal findings. We evaluate the performance of two state-of-the-art pre-trained models—ResNet-50 and Vision Transformer (ViT-L16)—and compare them against a baseline CNN. ResNet-50 achieves the highest performance, followed by ViT-L16, in both binary and multi-class tasks. Our analysis highlights each model’s accuracy, robustness, and generalization ability. The results demonstrate the potential of deep learning to improve diagnostic reliability and support scalable, consistent thoracic disease screening.