Noninvasive Prediction of Perineural Invasion and Lymphovascular Invasion in Prostate Cancer Using bpMRI Radiomic Signatures.
Authors
Affiliations (4)
Affiliations (4)
- The First Clinical Medical College of Lanzhou University, Lanzhou 73000, China (Y.F.Z., H.H., F.Z.).
- Department of Geriatric General Surgery, Sichuan Provincial People's Hospital, School of Medicine, University of Electronic Science and Technology of China, Chengdu 611731, China (C.Z.).
- The First Clinical Medical College of Gansu University of Chinese Medicine, Lanzhou 730000, China (D.L., H.C., Q.W., J.W., W.Z., X.W., Y.R., F.Z.).
- The First Clinical Medical College of Lanzhou University, Lanzhou 73000, China (Y.F.Z., H.H., F.Z.); The First Clinical Medical College of Gansu University of Chinese Medicine, Lanzhou 730000, China (D.L., H.C., Q.W., J.W., W.Z., X.W., Y.R., F.Z.); Department of Urology, Gansu Provincial Hospital, Lanzhou 730000, China (F.Z.). Electronic address: [email protected].
Abstract
Perineural invasion (PNI) and lymphovascular invasion (LVI) are critical predictors of aggressive behavior and poor prognosis in prostate cancer (PCa), yet their diagnosis relies on postoperative histopathology. This study aims to develop a noninvasive radiomic model based on biparametric magnetic resonance imaging (bpMRI) for preoperative prediction of PNI and LVI. A total of 256 patients with pathologically confirmed PCa who underwent radical prostatectomy were retrospectively enrolled. Patients from Center 1 (n = 179) constituted the training set, while those from Center 2 (n = 77) formed the external test set. A rigorous imaging-pathology correlation protocol was applied to ensure accurate lesion matching. Inter-observer variability in segmentation was assessed (ICC > 0.75 for 85% of features), with final ROIs determined by consensus. Radiomic features were extracted from T2-weighted and diffusion-weighted imaging. Feature selection was performed using Spearman's correlation and LASSO algorithm. Multiple machine learning classifiers were constructed and interpreted with SHAP. The best-performing model for PNI prediction was Multilayer Perceptron (MLP), with an AUC of 0.805 (95% CI: 0.741-0.869) in the training set and 0.795 (95% CI: 0.698-0.896) in the test set. For LVI prediction, Logistic Regression achieved the highest performance, with an AUC of 0.859 (95% CI: 0.804-0.914) in the training set and 0.810 (95% CI: 0.714-0.906) in the test set. Calibration curves and decision curve analysis indicated good model accuracy and clinical utility. Radiomic models derived from bpMRI can noninvasively and robustly predict PNI and LVI in PCa, demonstrating good generalizability across independent cohorts.