PURPOSE

To quantify population-specific differences in prostate cancer (PCa) presentation between African American (AA) and White (W) men on MRI using radiomics.

MATERIALS AND METHODS

We identified N = 149 men with PCa who underwent 3T MRI, a confirmatory biopsy and for whom self-reported race was available. Patient studies were partitioned into training (D) and hold-out test set (D). Three hundred radiomic features quantifying textural patterns were extracted from radiologist delineated PCa regions of interest (ROI) on biparametric MRI. Features with significant differences ( < .05) between clinically significant (csPCa) and insignificant (ciPCa) PCa were identified. Machine learning models were trained separately for AA and W men (C, C) on D to distinguish csPCa and ciPCa. Validation on D was assessed for AUC and compared against a population agnostic model (C) in combination with clinical parameters (age, PSA, Prostate Imaging Reporting and Diagnostic System and tumor volume).

RESULTS

Radiomic features from PCa ROIs on biparametric MRI associated with csPCa were observed to be different in AA compared with W men, especially in the peritumoral region. Population-specific radiomic models outperformed similarly trained C models (AUC = 0.84, 0.57 with C, C; < .05) in AA men on D. Similar findings were observed for W men (AUC = 0.71, 0.60 with C, C; < .05). Integrating clinical and radiomics further improved the risk stratification for AA men (AUC = 0.90) and W men (AUC = 0.75).

CONCLUSIONS

Accounting for population-specific differences in radiomics may enable improved PCa risk stratification at MRI among AA men compared with a population agnostic approach.