Related Experiment Videos
Enhancing Deep Learning Chest Disease Diagnosis through Adversarial Training for Robust and Reliable Medical Imaging
Ayah Fareed Al Omar1, Shadi A Aljawarneh1
1Jordan University of Science and Technology, Faculty of Computer and Information Technology-AI and Data Science, Irbid, Jordan.
Current Medicinal Chemistry
|July 29, 2026
Summary
This study enhances deep learning models for chest X-ray analysis using adversarial training and transfer learning. The improved models accurately classify diseases like COVID-19 and pneumonia while remaining robust against perturbations.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Healthcare
- Computer-Aided Diagnosis
Background:
- Chest X-ray (CXR) is crucial for diagnosing thoracic diseases.
- Automated CXR analysis faces challenges due to imaging variability and diverse disease presentations.
- Developing robust and accurate AI models for CXR interpretation is essential.
Purpose of the Study:
- To investigate the integration of adversarial training with transfer learning for improved deep learning model robustness in multi-class chest disease classification.
- To maintain high diagnostic accuracy while enhancing resistance to adversarial perturbations.
- To assess the interpretability of the developed models for clinical decision support.
Main Methods:
- Utilized a real-world dataset of chest X-ray images for classifying normal, COVID-19, pneumonia, and asthma.
- Fine-tuned pre-trained convolutional neural networks (ResNet50, InceptionV3, InceptionResNetV2) using transfer learning.
- Incorporated adversarial robustness via Neural Structured Learning (NSL) and analyzed interpretability using Grad-CAM and LIME.
Main Results:
- The ResNet50 model achieved a high testing accuracy of 98.35%.
- The model demonstrated enhanced robustness, maintaining stable performance under adversarial perturbations.
- Explainable AI (XAI) visualizations highlighted clinically relevant regions, offering insights into model decision-making.
Conclusions:
- Integrating adversarial training with transfer learning significantly improves the robustness of deep learning models for chest X-ray analysis.
- Enhanced robustness does not compromise diagnostic accuracy.
- XAI techniques improve model interpretability, supporting their potential as reliable clinical decision-support tools.