Goal: Use multiparametric MRI and basic clinical variables (age, PSA, PSA density, prostate volume) to estimate the probability of clinically significant prostate cancer (csPCa) and recommend whether biopsy is warranted, before histopathological confirmation is available.
Input modalities
- Multiparametric MRI (T1w, ADC maps, DWI)
- Clinical variables: age, PSA, PSA density, prostate volume
Output required
- Probability of csPCa
- Binary biopsy recommendation (yes / no)
- Structured reasoning trace referencing PI-RADS scores, lesion characteristics, zonal anatomy, and capsular contact
Ground truth: Histopathology-confirmed csPCa (ISUP Grade Group >= 2); negative cases confirmed by longitudinal PSA follow-up.
Dataset
| Split | Cases | Notes |
|---|---|---|
| Training | 75 | Radboudumc; real-world class distribution |
| Validation | 75 | Radboudumc; up to 5 submissions allowed |
| Test | 250 | 100 cases from Karolinska Institute (external); 1 submission allowed |
Input data (precomputed): Precomputed MRI features as .h5 files + clinical variables as .csv. Raw images are not provided for inference.
Expected output per case: JSON file (~5 KB) containing the probability score and structured reasoning trace.
Primary metric: AUROC
Provided tool: Validated MRI prostate zone segmentation