Deep Learning-Based Multi-Label Classification of Thoracic Diseases using EfficientNet

Abstract

This paper presents a deep learning framework for multi-label classification of thoracic diseases using chest X-ray images and the EfficientNet architecture. It tackles the challenge of diagnosing multiple conditions that look similar and often occur together. The framework uses transfer learning and fine-tunes several EfficientNet variants (B0-B4) with the NIH ChestX-ray14 dataset. It also employs focal loss and stochastic weight averaging to address class imbalance and improve the model's ability to generalize. This approach uses attention mechanisms and Grad-CAM to make the findings easier to explain, showing that predictions rely on important clinical features. The model achieved high ROC-AUC scores for more distinct conditions like Effusion (0.8545), Cardiomegaly (0.8003), and Pneumothorax (0.7820). However, it had more trouble classifying subtler issues like Infiltration, Nodule, and Pneumonia. The framework performs as well as or better than several established CNN models regarding accuracy and efficiency. It offers a strong and repeatable workflow for radiology applications and has potential for use in other diagnostic tasks. The study also notes limitations, such as class imbalance and the need for better clinical decision support, and suggests future improvements like better data augmentation techniques and uncertainty modeling.