Anatomical Prior-Guided Black-Box Attack Optimization for Prostate Tumor mpMRI Classification Models

Area of Science:

• Medical Imaging AI
• Machine Learning Security
• Oncology Diagnostics

Background:

• Deep learning models for prostate cancer diagnosis using multi-parametric MRI (mpMRI) are advanced but vulnerable to adversarial attacks.
• Adversarial attacks on medical images face an efficiency-imperceptibility dilemma due to anatomical details and multi-modal variations.
• Existing black-box attacks are either imprecise or computationally expensive, hindering clinical translation of AI models.

Purpose of the Study:

• To propose a novel black-box optimization method that enhances the efficiency and imperceptibility of adversarial attacks on medical AI.
• To address the limitations of existing adversarial attack strategies in the context of medical imaging.
• To develop a robust framework for evaluating the security of AI models in clinical settings.

Main Methods:

• Anatomical Prior-Guided black-box optimization integrating mpMRI Regions of Interest (ROI).
• Utilizing local finite difference probing for model sensitivity analysis to create a dynamic weight map.
• Constraining optimization to diagnostically critical high-frequency regions for targeted perturbations.

Main Results:

• Reduced query overhead by 40%-70% compared to existing methods.
• Decreased pixel modification count (L0 norm) by over 40%.
• Demonstrated effective generation of minute yet disruptive perturbations in prostate mpMRI datasets.

Conclusions:

• Anatomical priors play a crucial role in balancing attack imperceptibility and efficiency for medical AI.
• The proposed method offers a low-resource framework for evaluating the robustness of medical AI against adversarial attacks.
• This research provides a new paradigm for securing AI diagnostic tools in healthcare.