Extraprostatic extension (EPE) refers to prostate cancer that has spread beyond the outer capsule of the prostate gland into surrounding tissue. Accurately identifying EPE before surgery is critical because it directly influences whether surgeons can safely perform nerve-sparing radical prostatectomy, a technique designed to preserve the nerves responsible for erectile function.
The neurovascular bundles lie directly adjacent to the prostate capsule. If cancer has extended into this area, attempting nerve preservation risks leaving cancer cells behind -- a condition called positive surgical margins (PSM). Positive margins are strongly associated with disease recurrence after surgery, making accurate preoperative EPE detection essential for both oncological safety and patient quality of life.
Current standard imaging with 3 Tesla multiparametric MRI (3T mpMRI) is the recommended tool for local prostate cancer staging, but its ability to detect EPE is limited. Meta-analyses report high specificity (90%) but poor sensitivity (only 57%) for EPE detection, meaning roughly half of EPE cases are missed by conventional radiological review.
Conventional radiomic features used to predict EPE -- such as tumor contact length (TCL) and apparent diffusion coefficient (ADC) -- are measured in 2D and may not fully capture the three-dimensional geometry of how a tumor relates to the prostate capsule. This study aimed to improve EPE prediction by incorporating novel 3D geometric features and AI-driven machine learning models.
The study analyzed 994 lesions from 794 prostate cancer patients who underwent robot-assisted radical prostatectomy (RARP) at two Dutch hospitals: St. Antonius Hospital (SAZ) and Canisius Wilhelmina Hospital (CWZ). An independent external cohort from Radboud University Medical Center provided 189 additional lesions for external validation, enabling rigorous testing of how the models perform on entirely new data.
MRI scans were acquired using 3T scanners from Siemens and Philips. All lesions and the prostate gland were segmented using Quantib Prostate AI software, an FDA-approved commercial tool. The software automatically generated prostate outlines on T2-weighted images, which were then reviewed and corrected by a trained annotator under the supervision of experienced uro-radiologists.
A key innovation in the study was the generation of a 3D prostate capsule model from the segmentation data. This allowed researchers to compute novel geometric radiomic features including the tumor contact surface area (TCSA) and tumor contact volume (TCV) -- measurements of how much of the tumor's surface touches or is near the prostate capsule in three dimensions, rather than relying on a simple one-dimensional length measurement.
To maintain class balance given the unequal numbers of EPE and non-EPE lesions, the researchers applied SMOTE (Synthetic Minority Over-sampling Technique), which generates synthetic examples of the minority class (EPE lesions) to prevent the model from being biased toward the more common non-EPE outcome during training.
Three machine learning classifiers were developed and compared: Random Forest (RF), Extra Trees (ET), and Logistic Regression (LR). All three were trained on 80% of the internal cohort lesions and tested on a held-out 20% internal test set, as well as the fully independent external cohort. Clinical variables including PSA levels and ISUP biopsy grade were incorporated alongside radiomics features.
Feature selection was performed using different methods suited to each model type. For RF and ET, impurity-based feature importance scores were used to rank and select the most predictive radiomic variables. For LR, elastic net regression -- a regularization method that penalizes unimportant features -- was used to select the final feature set.
Model performance was evaluated across multiple dimensions: discrimination (AUC), calibration (agreement between predicted probability and actual outcome), and clinical net benefit using decision curve analysis. This multi-faceted evaluation ensured that the models would be assessed not just on raw accuracy but on their practical usefulness for clinical decision-making at different risk thresholds.
Model robustness was further tested by varying geometric radiomic variables by plus or minus 5%, simulating potential measurement inconsistencies. This stress test assessed whether small changes in segmentation would dramatically alter the model's predictions -- an important check before considering clinical deployment.
All three AI models achieved strong performance in predicting EPE at the lesion-specific level. AUC values ranged from 0.86 to 0.91 on both the internal and external test cohorts, indicating excellent discriminative ability. The Logistic Regression model achieved the highest AUC of 0.91 on the external cohort, while the Random Forest model achieved the best overall calibration.
Calibration analysis showed that the Random Forest model had the highest agreement between predicted and observed probabilities in both test cohorts, with overlapping confidence intervals indicating consistent performance across datasets. The LR model calibrated well on the internal cohort but underestimated risk above a 30% probability threshold. The ET model showed poor calibration in both cohorts.
Decision curve analysis (DCA) -- a method for evaluating the practical clinical benefit of a prediction model across a range of risk thresholds -- showed that all three models provided greater net benefit than either treating all patients (maximum nerve-sparing) or treating none (no nerve-sparing), within the clinically relevant risk range of 0 to 40%. The LR model showed the highest net benefit, closely followed by RF.
A direct comparison with experienced uro-radiologists using standard PI-RADS reporting revealed that the AI models significantly outperformed human radiological assessment. On the internal test cohort, the Random Forest model achieved significantly higher sensitivity (72% vs. 53%, p = 0.02) and negative predictive value (83% vs. 75%, p = 0.01) compared to radiologists, meaning it was better at detecting EPE when it was present.
On the external test cohort, the RF model demonstrated significantly better specificity (89% vs. 69%, p < 0.001) and positive predictive value (80% vs. 55%, p < 0.001) compared to radiological reporting. Overall accuracy was significantly higher for both the RF and LR models compared to radiology (83% vs. 67%, p = 0.02), a clinically meaningful 16-percentage-point improvement.
The feature importance analysis of the Random Forest model identified tumor contact length (TCL), lesion diameter, lesion volume, and the novel tumor contact surface area (TCSA) as the most important predictors of EPE. The inclusion of TCSA alongside conventional features was found to improve EPE prediction, demonstrating the added value of 3D geometric characterization.
The most immediate clinical application of this work is in guiding the extent of nerve-sparing surgery during radical prostatectomy. If surgeons know preoperatively which specific lesion locations have extended beyond the prostate, they can make more informed decisions about where to preserve or sacrifice the neurovascular bundles, potentially reducing both the risk of leaving cancer behind and the risk of unnecessary damage to erectile function.
The lesion-specific approach is a critical advancement over previous whole-gland or side-specific EPE models. In particular, anteriorly located tumors may be far from the neurovascular bundles even when EPE is present, meaning nerve-sparing can be safely performed on the opposite side. A lesion-specific model captures this spatial nuance, while coarser whole-gland approaches cannot.
The 3D geometric data generated by the AI segmentation pipeline also has potential for patient communication and shared decision-making. Visualizing where a tumor contacts the prostate capsule in three dimensions could help patients better understand their individual risk profile and the trade-offs involved in choosing between nerve-sparing and non-nerve-sparing surgery.
The authors caution that prospective data is still needed to establish whether improved preoperative EPE prediction actually translates into improved surgical outcomes such as reduced positive surgical margins and lower recurrence rates. Retrospective validation, while encouraging, cannot fully substitute for prospective clinical trials.
This study is the first to present an externally validated, lesion-specific radiomics-based model for EPE prediction that incorporates novel 3D geometric features such as tumor contact surface area (TCSA) and tumor contact volume (TCV). The Random Forest model demonstrated the best overall performance profile, combining strong discrimination, good calibration, and robustness to small measurement variations.
The AI models consistently and significantly outperformed conventional radiological reporting by experienced uro-radiologists, with the accuracy advantage holding up on an independent external dataset from a different institution. This external validation is particularly important as it demonstrates that the model is not just optimized for the training data but generalizes to new clinical settings.
Future research directions include incorporating information about the tumor's relationship with other nearby anatomical structures such as the urethra, bladder neck, and urinary sphincter, and eventually developing tools for real-time surgical navigation. Implementation of this approach in clinical practice will require prospective validation and integration into existing radiology and urology workflows.