Task 3: Prostatectomy Pathology Prediction¶

Goal: After radical prostatectomy, integrate whole-slide images from the prostatectomy specimen, biopsy findings, preoperative MRI, and preoperative serum PSA to estimate time-dependent biochemical recurrence (BCR) risk. Cases may include multiple MRIs and/or biopsies over time, and missing modalities are intentionally present.

Input data

• structured-prompt.json: structured clinical variables, such as age, preoperative PSA, and DRE findings
• prostate-time-to-recurrence-or-last-follow-up-clinical-data.json: radiology report, biopsy pathology report, surgical pathology report, prior clinical notes, and family history
• prostate-modality-level-neural-representations.json: MRI, biopsy WSI, and prostatectomy WSI neural representations

See Data Sources and Imaging for the complete input structure and variable descriptions.

Output required

• Predicted BCR status within the observation window (1 = BCR; 0 = no BCR)
• Free-text reasoning providing a clinical justification for the predicted outcome, e.g., "Based on xxx and taking xxx into consideration, this patient is not expected to develop biochemical recurrence within the next x months."

Ground truth: BCR defined as confirmed PSA rise >= 0.2 ng/mL post-prostatectomy. Patients without recurrence at last follow-up, as well as those who received salvage treatment, migrated, died, or were otherwise lost to follow-up, are treated as censored observations.

Expected output per case: Two JSON files (~5 KB) containing the predicted time to outcome, the event indicator, and the corresponding free-text reasoning.

Provided neural representations: MRI features from an MRI-based prostate cancer detection model, biopsy WSI features from an automated Gleason grading model, and prostatectomy WSI features extracted using a standardized preprocessing and feature-extraction pipeline.

Primary metric: Harrell's C-index