Improving PI-RADS Prediction and Prostate Cancer Detection from Multiparametric MRI with Anatomical Guidance
Organizations: Weill Cornell Medicine, New York, NY, USA · Cornell Tech, New York, NY, USA
Abstract
Accurate prediction of Prostate Imaging Reporting and Data System (PI-RADS) scores from multiparametric MRI (mpMRI) remains challenging, more so than clinically significant prostate cancer (csPCa) detection. Given the importance of lesion location in PI-RADS scoring, we evaluated whether incorporating prostate zonal masks improves PI-RADS prediction and csPCa detection across multiple architectures (3D ResNet-18, ResNet-50, and ViT) using a large-scale internal dataset. Anatomical guidance generally improved performance across architectures, with the best results achieved by ResNet-18 with 2-zone guidance for PI-RADS prediction (QWK = 0.511) and ResNet-50 with 5-zone guidance for csPCa detection (AUC = 0.809). Fine-tuned results on external datasets further demonstrated the benefits of anatomical guidance. Furthermore, models with anatomical guidance improved the utility of model-predicted PI-RADS scores for PI-RADS-based csPCa detection (positive defined as PI-RADS>=3). Combining model-predicted and expert-assigned PI-RADS scores significantly increased the AUC from 0.711 to 0.756 (p < 0.05), suggesting that AI-derived PI-RADS scores may provide complementary information to expert assessment for csPCa detection. Overall, our findings highlight the value of incorporating prostate zonal anatomy into deep learning-based prostate MRI assessment.
Figures & tables
| Dataset | Task | Train | Val | Test | Total |
|---|---|---|---|---|---|
| Internal | PI-RADS | 10,852 | 688 | 1,933 | 13,473 |
| csPCa (adjusted) | 6,491 | 437 | 1,257 | 8,185 | |
| UCSF | PI-RADS | 678 | 98 | 195 | 971 |
| csPCa | 680 | 98 | 195 | 973 | |
| PMUB | PI-RADS (pseudo) | 159 | 95 | 382 | 636 |
| csPCa | 167 | 100 | 401 | 668 |
| PI-RADS (Ordinal multi-label; 1, 2, 3, 4, 5) | csPCa (Binary-label; Positive/Negative) | |
| QWK | AUC | |
| ProFound-Conv | 0.181 (0.141, 0.217) | 0.710 (0.680, 0.741) |
| 3D ViT | 0.182 (0.140, 0.225) | 0.636 (0.604, 0.669) |
| 3D ViT+2zones | 0.215 (0.172, 0.257) | 0.678 (0.648, 0.708) ∗ |
| 3D ViT+5zones | 0.232 (0.190, 0.272) ∗ | 0.689 (0.659, 0.719) ∗ |
| 3D ResNet-18 | 0.471 (0.434, 0.507) | 0.789 (0.761, 0.816) |
| PI-RADS (Ordinal multi-label; 3, 4, 5) | csPCa (Binary-label; Positive/Negative) | ||
|---|---|---|---|
| QWK | AUC | ||
| UCSF | 3D ResNet-18, scratch | 0.071 (-0.045, 0.184) | 0.503 (0.416, 0.580) |
| 3D ResNet-18 | 0.152 (0.020, 0.280) | 0.633 (0.552, 0.706) | |
| 3D ResNet-18+2zones | 0.167 (0.032, 0.291) | 0.668 (0.586, 0.746) | |
| 3D ResNet-18+5zones | 0.125 (-0.007, 0.253) | 0.624 (0.541, 0.697) | |
| 3D ResNet-50, scratch | 0.053 (-0.064, 0.161) | 0.590 (0.509, 0.671) |
| AUC | |
|---|---|
| ProFound-Conv | 0.597 (0.547, 0.641) ∗ |
| ViT | 0.588 (0.539, 0.634) ∗ |
| ViT+2zones | 0.639 (0.593, 0.679) ∗ |
| ViT+5zones | 0.653 (0.607, 0.694) ∗ |
| ResNet-18 | 0.678 (0.637, 0.717) |
| ResNet-18+2zones | 0.710 (0.668, 0.750) |
Medical Image Analysis
Medical image analysis uses computational methods to interpret clinical images, measure anatomy and disease, and support diagnosis and treatment.
See all 309 papers