Robust Medical Diagnosis: A Novel Two-Phase Deep Learning Framework for Adversarial Proof Disease Detection in...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in diagnostics
• Deep learning applications

Background:

• Deep learning (DL) models excel at diagnosing diseases from radiology images, crucial during the COVID-19 pandemic.
• Despite their accuracy, DL models are vulnerable to adversarial attacks, which can compromise diagnostic reliability and patient care.

Purpose of the Study:

• To develop and evaluate a robust defense framework against adversarial attacks on deep learning models used for medical image diagnostics.
• To enhance the resilience and reliability of AI-powered diagnostic tools for conditions like COVID-19.

Main Methods:

• Proposed a two-phase defense framework incorporating adversarial learning during training and JPEG compression for filtering during inference.
• Evaluated the framework on ResNet-50, VGG-16, and Inception-V3 models for classifying lung radiology images (X-ray, CT) into normal, pneumonia, and COVID-19 pneumonia categories.
• Assessed model vulnerability against Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Basic Iterative Method (BIM) adversarial attacks.

Main Results:

• Adversarial attacks significantly degraded the performance of baseline DL models.
• The proposed defense framework substantially improved model resistance to adversarial attacks, maintaining high accuracy on perturbed data.
• The framework ensured the reliability of AI models in diagnosing COVID-19 from clean medical images.

Conclusions:

• The developed defense framework effectively mitigates adversarial attacks on deep learning diagnostic models.
• This approach enhances the trustworthiness of AI in medical diagnostics, particularly for critical applications like COVID-19 detection.
• Ensuring model robustness is vital for the safe and reliable deployment of AI in healthcare settings.