Improving Prostate Cancer (PCa) Classification Performance by Using Three-Player Minimax Game to Reduce Data Source...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer (PCa) classification is challenging due to diverse tissue patterns and data heterogeneity from varied collection methods.
• Existing generalized PCa classifiers struggle with inconsistent data, limiting their effectiveness.

Purpose of the Study:

• To introduce a novel Generative Adversarial Network (GAN)-based three-player minimax game framework to overcome data source heterogeneity in PCa classification.
• To enhance the performance of PCa classification using high-frequency ExactVu ultrasound data.

Main Methods:

• Utilized a GAN-based three-player minimax game framework for the first time in PCa classification.
• Employed a modified U-Net as the encoder within the GAN framework.
• Analyzed high-frequency ExactVu ultrasound data from 693 patients across five data centers, with Gleason Scores assigned to 12 prostatic regions.

Main Results:

• Achieved an Area Under the ROC Curve (AUC) of 93.4% for benign vs. malignant classification.
• Obtained a sensitivity of 95.1% and specificity of 87.7% for benign vs. malignant classification.
• Demonstrated significant improvements of 5.0% (AUC), 3.9% (sensitivity), and 6.0% (specificity) compared to models trained on heterogeneous data.

Conclusions:

• The proposed GAN-based three-player minimax game framework effectively tackles data source heterogeneity in PCa classification.
• The framework significantly improves PCa classification performance, particularly in distinguishing benign from malignant cases.
• This study highlights the potential of advanced AI techniques for more accurate and generalized cancer diagnostics.