ProViCNet

ProViCNet (Prostate Vision Contrastive Network) is an organ-specific vision foundation model for detecting and localizing clinically significant prostate cancer (csPCa) in radiology images. It was developed by the PIMed group at Stanford University with collaborators at the Catholic University of Korea and UC San Diego, released as a preprint in February 2025 and published in npj Digital Medicine in 2026.

The clinical problem is interpretive variability. Radiologists reading multiparametric MRI (mpMRI) miss roughly 12% of clinically significant cancers in biopsy cohorts, and reported specificity spans 21.9% to 68.5% depending on the diagnostic criteria applied. That combination sends many patients to unnecessary biopsy while still allowing aggressive disease to go undetected. General-purpose medical vision foundation models exist, but before ProViCNet there was no foundation model built specifically around prostate anatomy and the imaging modalities used in the prostate diagnostic pathway.

ProViCNet addresses this by adapting a general vision backbone to prostate imaging through label-guided, patch-level self-supervision, and by covering both mpMRI and transrectal ultrasound (TRUS) rather than MRI alone. Because ultrasound guides the biopsy itself, a single model that reads both modalities fits the real diagnostic workflow more closely than an MRI-only detector.

#Key Features

• Patch-level contrastive learning: Contrastive pairs are sampled at the image-patch level under biopsy-confirmed radiologist annotations, teaching the model to distinguish cancerous from normal tissue even near ambiguous lesion boundaries where whole-image objectives blur the distinction.
• Dual modality coverage: Separate pretrained models handle mpMRI (T2-weighted, diffusion-weighted imaging, and apparent diffusion coefficient maps) and TRUS, extending automated detection to the ultrasound guidance step of the biopsy pathway.
• Sequence-specific decoders: Each MRI sequence passes through a dedicated decoder producing its own probability map; those maps are then fused, so complementary anatomical detail from T2 and functional contrast from DWI/ADC both reach the final prediction.
• Lesion-level and sextant reporting: Predictions are evaluated by dividing the prostate into six regions, matching how urologists reason about location when planning targeted biopsy or focal therapy.
• Virtual screening test: Model output combined with serum PSA forms a biopsy-decision score that raises specificity while preserving sensitivity for aggressive disease.

#Technical Details

ProViCNet employs a 3D-enhanced vision transformer initialized from DINOv2 pretraining, coupled with the patch-level contrastive objective. Training and evaluation drew on 4,401 patients across six cohorts. Of these, 1,404 patients were split 80:20 for training and internal validation, with performance then measured on five held-out cohorts: C1 (n=352, biopsy-confirmed), C2 (n=120, radical prostatectomy with histopathology ground truth), C3 (n=1,497) and C4 (n=1,154) drawn from public external datasets, and C5 (n=292) from an independent external center.

Across those cohorts, area under the ROC curve for mpMRI ranged from 0.875 to 0.966. On C1 the model reached patient-level AUROC 0.923, sensitivity 0.895, and specificity 0.778; on the prostatectomy cohort C2, AUROC 0.875 with sensitivity 0.819 and specificity 0.730. Dice similarity coefficients for lesion segmentation were 0.425 and 0.389 on C1 and C2. In a reader study, ProViCNet outperformed radiologists (AUROC 0.907 versus 0.805, p<0.001). TRUS performance was lower, at 0.670 to 0.740, reflecting the harder signal characteristics of ultrasound. Combining model output with PSA raised screening specificity from 15% to 38% (p<0.001) while maintaining sensitivity for clinically significant cancer.

#Applications

ProViCNet supports MRI-based screening triage, targeted biopsy planning, and focal treatment planning. Its most immediate value is reducing negative biopsies: the PSA-combined screening score keeps sensitivity for aggressive disease while roughly halving the false-positive burden that drives unnecessary procedures. Because the released checkpoints run without site-specific retraining, radiology groups can apply them to their own mpMRI studies as a second reader, and researchers can use the pretrained encoder as a prostate-specific feature extractor for downstream tasks such as risk stratification or progression monitoring.

#Impact

ProViCNet is one of the first foundation models built around a single organ and its full imaging pathway rather than a single modality, and its cross-institution external validation — including two public datasets — is unusually thorough for a medical imaging model of its scale. Code and pretrained weights are released under an MIT license, which makes independent replication practical. Limitations remain: TRUS accuracy trails MRI substantially, segmentation Dice scores are modest even where detection AUROC is high, and all evaluation is retrospective, so prospective trials would be required before the virtual screening test could change biopsy practice.