Removing Adversarial Noise in X-ray Images via Total Variation Minimization and Patch-Based Regularization for Robust...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology
• Deep Learning for Disease Diagnosis

Background:

• Deep learning enhances radiology-based disease diagnosis accuracy using medical images.
• COVID-19 pandemic highlighted the need for rapid AI-driven diagnosis.
• Adversarial attacks pose a significant threat to deep learning model reliability in medical applications.

Purpose of the Study:

• To introduce a novel total variation minimization approach to mitigate adversarial noise in X-ray images for COVID-19 diagnosis.
• To evaluate the effectiveness of denoising techniques in improving the performance of AI models against adversarial attacks.

Main Methods:

• A classification model was developed using transfer learning for COVID-19 pneumonia detection from lung X-rays.
• The model's vulnerability to adversarial attacks was assessed using the Fast Gradient Sign Gradient (FGSM) method.
• A total variation minimization-based denoising approach was applied to adversarial X-ray images.

Main Results:

• Adversarial attacks reduced model accuracy from 95.56% to 19.83%.
• The proposed denoising method significantly improved accuracy on adversarial images, increasing it to 88.23%.
• The denoising approach demonstrated high efficacy in restoring diagnostic performance.

Conclusions:

• Denoising techniques can effectively enhance the resilience of AI-based COVID-19 diagnostic systems against adversarial attacks.
• The study provides a foundation for deploying robust AI diagnostic tools in clinical settings.
• Total variation minimization is a promising strategy for securing deep learning models in medical imaging.