MRI-based morphologic parameter model for preoperative prediction of transurethral thulium laser enucleation of the prostate difficulty.

Background: Benign prostatic hyperplasia (BPH) is associated with multiple long-term urinary complications, including obstructive renal failure if left untreated. Transurethral thulium laser enucleation of the prostate (ThuLEP) has become an increasingly popular method in the treatment of BPH, as it has reduced postoperative bleeding compared to surgery. However, it requires correct plane removal and capsular integrity maintenance, resulting in a steep learning curve. Thus, identifying BPH patients for whom ThuLEP is feasible, despite the increased difficulty, may aid in optimizing their care. In this study, we established a predictive model that combined magnetic resonance imaging (MRI) features incorporated into a machine learning-based algorithm with specific clinical characteristics to quantitatively assess ThuLEP difficulty for BPH patients. Methods: The data of 278 BPH patients who underwent ThuLEP at the First Hospital of Shanxi Medical University between November 2023 and May 2025 were retrospectively collected. The patients were divided into training [152], testing [66], and validation [60] dataset groups. All the patients underwent prostate MRI. Prostate volume (PV), intravesical prostatic protrusion (IPP), attached mural nodules, and prostate morphological angles in four directions (i.e., superior, inferior, left-lateral, and right-lateral) were measured under 3D Slicer. These MRI imaging features were incorporated into seven machine learning algorithms [Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XG Boost), and Light Gradient Boosting Machine (LightGBM)] to establish predictive models for ThuLEP difficulty. The clinical characteristics associated with ThuLEP difficulty were also identified by univariate and multivariate LR analyses, resulting in a clinical model comprising age, PV, and creatinine and preoperative total prostate-specific antigen (tPSA) levels. The imaging and clinical models were combined to form a joint model. Model accuracy, generalizability, performance, and clinical utility were assessed using receiver operating characteristic (ROC) curves, SHapley Additive exPlanations (SHAP), confusion matrices, and decision curve analyses (DCAs), respectively, in all three patient groups. Results: Of the seven algorithms, LightGBM had the best results for the imaging model, with area under the curve (AUC) values of 0.96 [95% confidence interval (CI): 0.93-0.99] and 0.91 (95% CI: 0.83-0.98) for the training and testing datasets, respectively. Moreover, the joint model that combined the imaging and clinical models had the highest accuracy for determining ThuLEP difficulty, with AUC values of 0.967 (95% CI: 0.940-0.986), 0.924 (95% CI: 0.852-0.978), and 0.930 (95% CI: 0.870-0.979) for the training, testing, and validation datasets, respectively. In comparison, the imaging model had AUCs of 0.945 (95% CI: 0.905-0.977), 0.895 (95% CI: 0.812-0.961), and 0.875 (95% CI: 0.790-0.952), while the clinical model had AUCs of 0.919 (95% CI: 0.871-0.955), 0.880 (95% CI: 0.794-0.950), and 0.893 (95% CI: 0.796-0.971) for the same datasets. The joint model also had greater clinical utility in the DCA compared to the other two models, as well as high values for true positives and negatives in the confusion matrices. Conclusions: The six MRI imaging features in the LightGBM algorithm, combined with the four clinical characteristics, accurately predicted ThuLEP difficulty. Our model could aid in the development of safe personalized treatment approaches for BPH patients.