Anatomical Context Protects Deep Learning from Adversarial Perturbations in Medical Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neuroscience

Background:

• Deep learning models excel in medical image processing tasks.
• However, deep neural networks are vulnerable to subtle adversarial perturbations.
• These vulnerabilities pose risks in clinical applications.

Purpose of the Study:

• To investigate the impact of adversarial perturbations on age prediction from 3D MRI brain images.
• To compare the robustness of a conventional deep neural network against a hybrid deep learning model.
• To assess the effectiveness of anatomical context features in mitigating adversarial attacks.

Main Methods:

• Utilized 3D MRI brain images for age prediction.
• Developed and evaluated a conventional deep neural network.
• Implemented and tested a hybrid deep learning model incorporating anatomical features.
• Introduced imperceptible adversarial perturbations to input images.

Main Results:

• Adversarial perturbations significantly increased errors in predicted age.
• A single perturbation could affect large batches of images.
• The hybrid model demonstrated substantially higher robustness to adversarial attacks compared to the conventional model.

Conclusions:

• Deep learning models for medical image analysis are susceptible to adversarial attacks.
• Hybrid models integrating anatomical context offer improved resilience.
• Further research is needed to enhance the security and reliability of AI in clinical settings.