Automated prostate segmentation can streamline prostate MRI analysis, enabling objective volume measurements, automated PSA density calculation and extraction of quantitative imaging biomarkers. However, achieving reliable performance across scanners, acquisition protocols and patient populations remains a major challenge.

This study developed a deep learning model for automated multi-regional segmentation of T2-weighted prostate MRI, simultaneously identifying the whole prostate gland, central-transition zone (CZ-TZ), peripheral zone (PZ) and seminal vesicles (SV).

A key strength is its highly heterogeneous, intercontinental dataset: 243 MRI examinations from 7 countries across 3 continents, acquired on 10 scanner models from three major vendors. This represents a substantially broader real-world validation setting than previous prostate segmentation studies.

The model achieved Dice scores of 0.88 for the whole prostate, 0.85 for CZ-TZ, and 0.72 for both PZ and seminal vesicles, with no significant performance differences across manufacturers or between the independent US cohort and other test cases.

These findings demonstrate the potential of generalizable AI-powered prostate MRI segmentation to support PI-RADS-based assessment, quantitative imaging biomarkers and more efficient prostate cancer workflows across heterogeneous clinical environments.