T6: Detecting Clinically Significant Cancer in Prostate MRI Exams¶
Objective:
Detect clinically significant prostate cancer (csPCa) lesions using 3D prostate MRI scans. Clinically significant prostate cancer is defined as lesions with an ISUP score ≥ 2.
Patient Population:
- Patients suspected of having clinically significant prostate cancer (e.g., elevated PSA or abnormal DRE findings).
- No prior prostate-specific treatments.
- No previous diagnosis of ISUP ≥ 2 cancer.
Imaging Data:
- Axial T2-weighted (T2W) MRI
- Axial high b-value diffusion-weighted imaging (DWI, b ≥ 1000 s/mm²)
- Axial apparent diffusion coefficient (ADC) maps
Test Data:
- 1000 cases, from three Dutch centers and one Norwegian center.
- Scans acquired using Siemens or Philips MRI scanners.
Reference Standard:
- Positive cases: Confirmed histopathological evidence (ISUP ≥ 2).
- Negative cases: ISUP ≤ 1 or MRI PI-RADS ≤ 2, supported by ≥ 3 years of clinical follow-up.
- Voxel-level annotations created using ITK-SNAP v3.80 by trained investigators or a radiology resident, supervised by expert radiologists.
- All annotations underwent rigorous quality control at the coordinating center (RUMC).
Evaluation Metrics:
Performance is assessed using the unweighted average of:
- Area Under the Receiver Operator Characteristic Curve (AUROC)
- Area Under the Precision-Recall Curve (AP)
Relation to Existing Challenges:¶
- Task 6 is adapted from the PI-CAI challenge.
- Validation and test datasets mirror those used in PI-CAI but omit clinical parameters and coronal/sagittal T2 images.
- PI-CAI aimed to train task-specific models ("narrow AI") on large, partially labeled datasets, whereas UNICORN emphasizes adaptation through few-shot learning on the challenge platform.