Multi-structure prostate MRI segmentation underpins biopsy guidance, radiotherapy planning, and morphometric reporting. Most deep-learning tools published to date have tackled one structure at a time, have rarely been tested on data from institutions they never saw during training, and have seldom converted their segmentation masks into measurements a clinician can actually read. A new retrospective multi-center study, published in BMC Medical Imaging, reports a framework designed to close all three gaps at once.
The model, named NineSeg-Prostate, is built on a 2D MedNeXt network and was trained to jointly segment nine structures on T2-weighted MRI: the four prostate zones, the lesion, and four adjacent pelvic organs—the rectum, seminal vesicle, bladder, and urethra. The training data came from 317 patients contributing 6,399 slices across three tertiary centers, with images acquired between April 2020 and August 2022. An independent external cohort of 65 patients from the public Prostate-X dataset served as the out-of-distribution test. The authors benchmarked their network against two established competitors, nnU-Net 3D and VGG-AG-UNet, and evaluated every output against expert annotations.
Performance was quantified with the Dice similarity coefficient, a standard overlap metric, together with the 95% Hausdorff distance for boundary accuracy. On the internal evaluation, MedNeXt outperformed both comparators on the peripheral zone (Dice 0.81), the transition zone (0.80), and the lesion (0.63). The network is also unusually lean: 10.7 million parameters and an inference speed of 26.4 frames per second.
Generalization was not uniform across structures but followed a clear gradient. External validation showed minimal degradation for the rectum and bladder (ΔDSC − 1.1% to − 4.5%), while the lesion lost 19.1% of its Dice score.
Beyond masks, the framework converts segmentations into six linear morphometric measurements—urethral length, lesion maximum diameter, and four zonal diameters—and compares them with the same measurements derived from expert annotations. Agreement was statistically significant for all six parameters, with intraclass correlation coefficients of at least 0.60 and P values below 0.001 throughout. The urethra length showed a mean absolute error of just 0.19 centimeters. Volumetric outputs were assessed with the same rigor, using mean absolute and relative error alongside the correlation metrics.
Three radiologists then independently rated the clinical usability of the automated contours on a 0–2 scale. Most structures received the top rating of 2, which the authors interpret as support for a radiologist-supervised workflow in which the model drafts the annotation and the human verifies and corrects it.
The study also prespecified a set of clinical correlative analyses. Lesion Dice scores were stratified by lesion volume and patient age. Zonal and lesion volumes, along with the morphometric indices, were compared across strata of ISUP grade, TNM stage, and PI-RADS score. Finally, Spearman correlations were computed between six clinical variables—age, body mass index, total PSA, Gleason grade, TNM stage, and PI-RADS—and twelve imaging measurements derived from the automated masks.
Lesion size, whether measured as volume, maximum diameter, or lesion-to-prostate ratio, correlated significantly and negatively with Gleason/ISUP grade, with all P values below 0.001, and lesion size also correlated with TNM stage, with correlation coefficients of at least 0.24.
The authors conclude that the 2D MedNeXt framework efficiently segments all nine prostatic and periprostatic structures, automates MRI-based morphometry, and demonstrates structure-dependent external generalizability, with potential utility as a first-pass annotation tool.
Subject of Research: Deep-learning segmentation of nine prostatic and periprostatic structures on prostate MRI with external validation and morphometric-clinical correlation
Article Title: Deep learning-based multi-center segmentation of nine prostatic and periprostatic structures: external validation and morphometric–clinical correlation
Article References: Xu, S., Jiang, L., Fan, R., Cao, X., Li, T., Liu, Z., Li, L., Zhang, F., Zhou, X., Wang, Y., & Wu, Y. (2026). Deep learning-based multi-center segmentation of nine prostatic and periprostatic structures: external validation and morphometric–clinical correlation. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02943-7
Image Credits: AI Generated
DOI: 10.1186/s12880-026-02943-7
Keywords: prostate MRI, deep learning, medical image segmentation, MedNeXt, external validation, morphometry, PI-RADS, Gleason grade, radiotherapy planning, Prostate-X, Dice similarity coefficient, clinical usability
News Source: Blake Davidson. (November 8, 2026). Research report: Deep learning-based multi-center segmentation of nine prostatic and periprostatic structures: external validation and morphometric–clinical correlation. Scienmag.



