A Multiparametric MRI-based Model for decoding Extraprostatic Extension in Prostate Cancer via Habitat-guided Radiomics and Clinical Integration.

[RATIONALE AND OBJECTIVES] Develop and validate a multiparametric MRI-based integrative model that combines intratumoral microenvironmental heterogeneity, peritumoral radiomic features, and clinicopathological variables for the noninvasive preoperative prediction of extraprostatic extension (EPE) in prostate cancer, aiming to support surgical decision-making and reduce postoperative biochemical recurrence.

[METHODS] MRI and clinicopathological data from 590 prostate cancer (PCa) patients across four centers (August 2013-September 2023) were retrospectively collected and divided into a training cohort (n = 249), internal validation cohort (n = 106), and two external test cohorts (n₁ = 199, n₂ = 36). Radiomic and habitat features were extracted from T2WI, DWI, ADC, and DCE MRI sequences. Six models were constructed: Clinical, Radiomics, Peri2mm, Peri4mm, Peri6mm, and Habitat. These were subsequently integrated into a fusion model (Habitat + Peri6mm + Clinical). Model performance was evaluated using ROC analysis, DCA, calibration curve, DeLong's test, nomogram and SHAP interpretation.

[RESULTS] The habitat model outperformed all unimodal models, achieving AUCs of 0.978 (training), 0.893 (validation), 0.832 and 0.836 (external test sets). Among radiomics-based models, the Peri6mm model showed the highest performance, with AUCs ranging from 0.949 to 0.752. The fusion model achieved the best overall predictive performance, with AUCs of 0.982 (training), 0.921 (validation), and 0.853 and 0.839 (external).

[CONCLUSION] Habitat and peritumoral radiomic features were independently predictive of EPE. The fusion model, integrating clinical, peritumoral, and habitat-derived features, further enhanced preoperative predictive accuracy.