Deep learning-based multi-center segmentation of nine prostatic and periprostatic structures: external validation and morphometric–clinical correlation

Multi-structure prostate MRI segmentation underpins biopsy guidance, radiotherapy planning, and morphometric reporting, yet most deep-learning approaches address single structures in isolation, rarely undergo external validation, and seldom translate masks into interpretable morphometric measurements. We aimed to develop and externally validate a deep-learning framework for joint segmentation of nine prostatic and periprostatic structures on T2-weighted MRI, and to assess its morphometric agreement and clinical usability. This retrospective multi-center study enrolled 317 patients (6,399 slices) from three tertiary centers (April 2020–August 2022) and an independent external cohort (Prostate-X, n = 65). A 2D MedNeXt network (NineSeg-Prostate) was trained to jointly segment four prostate zones, the lesion, and four adjacent pelvic organs on T2-weighted MRI, and benchmarked against nnU-Net 3D and VGG-AG-UNet. Segmentation, volumetric, and six linear morphometric measurements were evaluated against expert annotations using Dice similarity coefficient (DSC), 95% Hausdorff distance, intraclass correlation coefficient, and mean absolute/relative error. Three radiologists independently rated clinical usability on a 0–2 scale. Prespecified clinical correlative analyses (i) stratified lesion DSC by lesion volume and patient age, (ii) compared zonal/lesion volumetric and morphometric indices across ISUP grade, TNM stage, and PI-RADS strata, and (iii) computed Spearman correlations between six clinical variables (age, body mass index, total PSA, Gleason grade, TNM stage, PI-RADS) and twelve imaging measurements. MedNeXt was compact and fast (10.7 M parameters, 26.4 frames per second [fps]) and outperformed comparators on peripheral zone (DSC 0.81), transition zone (0.80), and lesion (0.63). External validation revealed a structure-specific generalization gradient, with minimal degradation for rectum and bladder (ΔDSC − 1.1% to − 4.5%) but substantial loss for the lesion (− 19.1%). Morphometric agreement was significant for all six linear parameters (all P < 0.001; urethra length mean absolute error 0.19 cm). Most structures received the highest radiologist usability rating. Lesion size correlated significantly with Gleason grade and TNM stage (all r ≥ 0.24), whereas PI-RADS showed no significant between-group differences in morphometric measures. The proposed 2D MedNeXt framework efficiently segments nine prostatic and periprostatic structures and automates MRI-based morphometry, demonstrating structure-dependent external generalizability and potential utility as a first-pass annotation tool. NineSeg-Prostate model jointly segments four prostate zones (peripheral, transition, central, and anterior fibromuscular stroma), the lesion, and four adjacent pelvic organs (rectum, seminal vesicle, bladder, urethra) on T2-weighted MRI with only 10.7 M parameters (26.4 fps). Multi-center training across three clinical centers and external validation on Prostate-X (n = 65) revealed a structure-specific generalization gradient, with bladder/rectum stable (ΔDSC − 1.1% to − 4.5%). Six linear morphometric indices—urethral length, lesion maximum diameter, and four zonal diameters—showed significant agreement with expert segmentations (all ICC ≥ 0.60, all P < 0.001). Lesion volume, lesion maximum diameter, and lesion-to-prostate ratio showed significant negative correlations with Gleason/ISUP grade (all P < 0.001); TNM staging was associated with lesion size and zonal volumes; PI-RADS scores showed no significant morphometric differentiation (all P > 0.05). Three radiologists rated most structures with the top usability score (score 2), supporting the feasibility of radiologist-supervised clinical evaluation.