The study cohort comprised men enrolled in active surveillance for prostate cancer (grade group 1 or favorable grade group 2) who had both fusion-targeted and systematic biopsies. Investigators performed automated gland segmentation using nnU-Net, reporting a mean Dice score of 0.89 ± 0.04 for the gland masks. That high Dice indicates consistent automated contouring across the dataset, according to the authors. Radiomic features were then extracted from those gland segmentations. The source did not report the full list of individual radiomic features retained after selection; it only specified that features were chosen using LASSO-penalized survival regression.
The selected radiomic variables were used to train an accelerated failure time model implemented with XGBoost. The study team contrasted this gland-level radiomics approach with two comparators: a model built on clinicopathologic variables and a model using lesion-level radiomics. The manuscript is a retrospective single-institution analysis; the authors allocated 410 men to model development and 112 to internal testing by site. The paper notes that the work is a preprint and has not been peer-reviewed.
Model discrimination favored the gland radiomics approach. Investigators reported a concordance index (C-index) of 0.77 for the gland radiomics model. By comparison, both the clinicopathologic model and the lesion radiomics model had C-indices of 0.68; the authors stated the differences were statistically significant after adjustment (adjusted P ≤ .017). Time-dependent area under the receiver operating characteristic curve (AUC) peaked at 36 months for the gland model, with a reported AUC of 0.87 (95% CI: 0.78, 0.94).
The team used Kaplan–Meier curves to compare progression-free survival across risk strata defined by the gland radiomics output. They also performed decision curve analysis to evaluate clinical net benefit, though the source text does not provide detailed decision curve plots or threshold-specific net benefit values beyond the reported risk-threshold results. The source did not disclose model hyperparameters, exact radiomic feature names, or details of internal cross-validation folds beyond the site-based split.
Investigators divided patients into high- and low-risk groups based on the gland radiomics model predictions. Progression-free survival at 36 months was 63.9% in the high-risk group versus 100% in the low-risk group. At 48 months the figures were 55.4% (high-risk) versus 93.8% (low-risk); log-rank P < .001 was reported for the between-group comparison. Those survival differences indicate separation in event timing between groups in the internal test set.
Using a risk threshold of 0.25, the model assigned 49.1% of patients as low risk at 36 months and 31.3% as low risk at 48 months. The negative predictive value at that threshold was reported as 1.00, while the positive predictive value was 0.36. The manuscript provides these operating characteristics but does not report sensitivity, specificity, or calibration statistics in the main abstract; additional metrics may be present in the full text.
This work is a retrospective single-institution study conducted under Institutional Review Board approval from the University of California, Los Angeles, as stated by the authors. The dataset included men imaged and biopsied from 2010 through 2025. The MRI and clinical data are not publicly available because of patient privacy and institutional restrictions; the authors said de-identified data may be shared on reasonable request and with IRB approval.
Limitations that the source explicitly reports include the single-site, retrospective design and the fact that the article is a preprint without peer review. The source does not report external validation in independent institutions or prospective testing. The authors did not provide a detailed list of radiomic features retained after LASSO or the exact model hyperparameters in the abstract; readers must consult the full manuscript or data/code links for those specifics.
Next steps and unresolved questions recorded in the source include external validation and broader assessment of clinical utility in prospective cohorts. The manuscript’s status as a preprint means clinical application would require peer review, reproducibility checks, and likely multi-institutional validation before adoption in routine active surveillance workflows.