Protecting Prostate Cancer Classification From Rectal Artifacts via Targeted Adversarial Training

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep neural networks (DNNs) are used for prostate cancer (PCa) classification using MRI.
• Rectal artifacts in MRI scans can significantly impair the accuracy of PCa classification by DNNs.
• Current DNN methods often fail to account for or mitigate the impact of rectal artifacts.

Purpose of the Study:

• To develop a novel strategy to enhance the robustness of DNN-based PCa classification against rectal artifacts.
• To introduce a method that specifically addresses the interference of rectal artifacts in prostate MRI analysis.

Main Methods:

• Proposed a Targeted adversarial training with Proprietary Adversarial Samples (TPAS) strategy.
• Generated proprietary adversarial samples incorporating rectal artifact patterns to challenge classification models.
• Trained PCa classification models using both original and generated adversarial samples.

Main Results:

• The TPAS strategy significantly improved PCa classification performance compared to ordinary training.
• Effectiveness was demonstrated across patient, slice, and lesion levels for both single- and multi-parametric MRI.
• Substantial performance gains were observed in recent advanced deep learning models.

Conclusions:

• The TPAS strategy is a valuable approach for mitigating the negative influence of rectal artifacts on deep learning models for PCa classification.
• This method enhances the reliability of AI-driven diagnostic tools in the presence of common imaging interferences.