Learning to predict prostate cancer recurrence from tissue images.

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Biochemical cancer recurrence (BCR) affects approximately 30% of men post-prostatectomy.
• Accurate prediction of BCR is crucial for timely intervention with surveillance or adjuvant therapy.
• Current prediction methods for BCR lack sufficient efficacy.

Purpose of the Study:

• To develop and evaluate PathCLR, a novel semi-supervised method for predicting prostate cancer recurrence.
• To utilize hematoxylin and eosin (H&E)-stained tissue microarrays (TMAs) for predicting BCR within five years.
• To assess PathCLR's performance against models based solely on clinicopathological features.

Main Methods:

• PathCLR employs a two-step process: self-supervised learning for feature representation, followed by a supervised neural network classifier.
• Training and validation were performed on two large prostate cancer datasets (CPCTR and JHU).
• Performance was evaluated using F1 scores and compared to models using only clinicopathological data.

Main Results:

• PathCLR achieved F1 scores of 0.61 (CPCTR) and 0.85 (JHU) using 10-fold cross-validation.
• PathCLR demonstrated statistically superior performance compared to models relying solely on clinicopathological features (P < .05).
• The method effectively integrates information from tissue images and clinicopathological data for improved prediction.

Conclusions:

• PathCLR offers a significant advancement in predicting prostate cancer recurrence.
• Tissue image analysis provides critical predictive information beyond traditional clinicopathological features.
• This approach holds promise for identifying patients who will benefit from intensified monitoring or treatment post-surgery.