Enhancing Deep Learning Chest Disease Diagnosis through Adversarial Training for Robust and Reliable Medical Imaging

Chest X-ray imaging is widely used for the diagnosis of thoracic diseases; however, developing reliable automated systems remains challenging due to variability in imaging conditions and disease presentation.

This study investigates whether integrating adversarial training with transfer learning can improve the robustness of deep learning models for multi-class chest disease classification while maintaining high diagnostic accuracy.

This study was approved by the Institutional Review Board (IRB)/Ethics Committee of King Abdullah II Hospital. A real-world dataset of chest X-ray images collected from King Abdullah University Hospital (KAUH) was used to classify four categories: normal, COVID-19, pneumonia, and clinically diagnosed asthma. Pre-trained convolutional neural networks, including ResNet50, InceptionV3, and InceptionRes- NetV2, were fine-tuned using transfer learning. Adversarial robustness was incorporated using Neural Structured Learning (NSL), and model interpretability was analysed using Grad-CAM and LIME.

The ResNet50-based model achieved the highest performance, with a testing accuracy of 98.35%. When evaluated under adversarial perturbations, the model maintained stable performance with only a slight decrease in accuracy, indicating improved robustness. XAI visualisations provided meaningful insights into model decision-making by highlighting clinically relevant regions in chest X-ray images.

The findings demonstrate that integrating adversarial training with transfer learning can enhance model robustness without compromising accuracy. In addition, the use of XAI techniques improves interpretability, supporting the potential use of such models as reliable clinical decision-support tools.