T3: Predicting the time to biochemical recurrence in H&E-stained prostatectomies¶

Objective:
Develop a model that predicts the time to biochemical recurrence (BCR) following radical prostatectomy. The model should output a continuous prediction for the time-to-event (BCR) or time-to-censoring (in years).

Patient Population:
149 patients treated at Radboudumc between 1992 and 2012 who underwent radical prostatectomy.

Imaging Data:
Each patient has one H&E-stained whole slide image (.tif) representing the prostate cross-section with the highest tumor grade (as per pathology report).

• Resolution: 0.5 microns per pixel
• Accompanied by a tissue segmentation mask (.tif):
  • Label 0: Background
  • Label 1: Tissue region
• For few shots, an event.json file with:
  • 0 if patient is censored (no BCR event)
  • 1 if patient is not censored (BCR occurred)

Test Data:
Unlabeled .tif whole slide images and tissue masks in the same format. Models must predict a float value for the time-to-event (BCR) or time-to-censoring (in years).

Reference Standard:

• Biochemical recurrence status (0 or 1), determined by rising PSA levels post-surgery
• Time to event (BCR) or censoring (last follow-up)
• No human annotation or Gleason pattern segmentation was used

Evaluation Metrics:
Model performance will be evaluated using the censored concordance index (c-index), which accounts for censored time-to-event data.

Relation to Existing Challenges:¶

• Task 3 is adapted from the LEOPARD challenge.
• Unlike LEOPARD, which relied on local training with large datasets, this task follows a few-shot learning paradigm, with training performed on-platform.