Adversarial training for prostate cancer classification using magnetic resonance imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning models are increasingly used for prostate cancer diagnosis.
• Improving the generalizability and diagnostic accuracy of these models is crucial.

Purpose of the Study:

• To investigate the effectiveness of adversarial training in enhancing deep learning models for prostate cancer diagnosis.
• To compare the performance of models before and after adversarial training.

Main Methods:

• Retrospective multicenter study including 396 prostate cancer patients.
• Development of four deep learning models (two for classification, two for grading) using MRI data.
• Retraining models with adversarial examples and comparing diagnostic performance using AUC and kappa scores.

Main Results:

• Adversarially trained models (AM1, AM2) showed significantly higher AUCs than baseline models (PM1, PM2) for prostate cancer classification in both internal and test datasets.
• Adversarially trained grading models (AM3, AM4) achieved higher kappa values than baseline models (PM3, PM4) across datasets.

Conclusions:

• Adversarial training effectively improves the generalizability and classification abilities of deep learning models for prostate cancer.
• This approach holds potential for advancing AI-driven diagnostic tools in oncology.