Adversarial attacks in radiology - A systematic review.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Cybersecurity in Healthcare
• Deep Learning Applications in Radiology

Background:

• Deep learning is increasingly used in radiology, raising cybersecurity concerns, especially regarding adversarial attacks.
• Adversarial attacks can manipulate AI models, potentially impacting diagnostic accuracy and patient safety.
• A systematic review is needed to understand the landscape of adversarial attacks in the field.

Approach:

• A systematic literature search was conducted using MEDLINE and Google Scholar up to April 2023.
• The review included 22 studies published between March 2018 and April 2023.
• Studies focused on adversarial attacks against various radiological image classification algorithms.

Key Points:

• Most studies (14/22) evaluated white-box attacks, with some achieving 100% success rates and reducing algorithm performance (AUC) to 0.
• Chest X-ray classification algorithms were the most frequent targets (11/22), followed by chest CT, brain MRI, and mammography.
• Attacks demonstrated significant potential to compromise the integrity of deep learning models in radiology.

Conclusions:

• Adversarial attacks represent a significant, albeit currently theoretical, risk to the safe deployment of deep learning in radiology.
• Proactive measures, including strengthened cybersecurity protocols and the development of ethical and legal frameworks, are crucial.
• Ensuring the secure and reliable use of AI in medical imaging requires ongoing vigilance and adaptation.