Safe Breast Cancer Diagnosis Resilient to Mammographic Adversarial Samples

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning models are vulnerable to adversarial data, leading to potential malfunctions.
• Ensuring AI accuracy and resilience is crucial for safe clinical applications.
• Breast cancer diagnosis using mammography requires robust AI systems resistant to adversarial attacks.

Purpose of the Study:

• To develop a novel method for building safe and robust deep learning models for breast cancer diagnosis.
• To enhance the resilience of AI models against adversarial samples in mammography.
• To improve the accuracy of AI-driven breast cancer diagnosis in the presence of adversarial data.

Main Methods:

• Proposed a novel adversarially robust feature learning (ARFL) method.
• Incorporated a feature correlation measure as an objective function to encourage robust features.
• Facilitated adversarial training using both standard and adversarial mammographic data.

Main Results:

• Evaluated deep learning diagnosis models on two independent clinical datasets (9,548 mammograms).
• Demonstrated that the ARFL method outperformed several state-of-the-art adversarial training techniques.
• Showcased ARFL's efficacy in enhancing adversarial training for robust breast cancer diagnosis.

Conclusions:

• ARFL is an effective method for enhancing adversarial training in medical AI.
• The proposed method contributes to building safer breast cancer diagnosis systems against adversarial attacks.
• This research paves the way for more secure AI applications in clinical settings.