Prostate cancer is one of the most common cancers among men worldwide, and nearly one in three patients treated surgically goes on to experience biochemical recurrence. The tools used to estimate that risk today, such as CAPRA-S, work well across a population but are much less reliable for an individual patient, partly because they depend on grading that varies between readers.
What the challenge asked
LEOPARD (LEarning biOchemical Prostate cAncer Recurrence from histopathology sliDes) asked whether a model can predict time to recurrence directly from the H&E slide, with no clinical variables at all. It ran from April to September 2024 and drew more than 440 participants from over 40 countries and more than 200 submissions. Sixteen teams submitted final models and nine went through to external validation. PAICON took part, and Witali Aswolinskiy and Christian Aichmüller are among the consortium authors.
The benchmark used 2,181 prostatectomy cases from four cohorts on three continents: Radboud University Medical Center in the Netherlands, the PLCO screening trial in the USA, Instituto Mario Penna in Brazil and University Hospital Cologne in Germany. Models were developed on the Radboud data, tuned on a small held-out set, then evaluated blind on 823 internal cases and finally on the two external cohorts.
What the models achieved
Most entries combined a pathology foundation model with multiple instance learning. The top five algorithms were pooled into an ensemble, which reached an overall C-index of 0.740, effectively level with histopathological grading at 0.739 and below CAPRA-S at 0.785. Performance held up across borders: 0.773 on the Brazilian cohort and 0.702 on the German one.
The more interesting result is what happens when the two are combined. Adding the AI ensemble to histopathological grading raised the C-index to 0.766, and adding it to CAPRA-S raised it to 0.799, a significant improvement over CAPRA-S alone.
Why it matters
Reading recurrence risk from the slide alone is already as informative as grading it by eye, and it is reproducible. More to the point, the signal the models pick up is not the signal the clinical tools already use: it adds to them rather than repeating them, which is what makes it worth putting in front of a clinician.