Biparametric MRI-Based Habitat Analysis Integrated With Deep Learning for Predicting Clinically Significant Prostate Cancer in PI-RADS Category 3 Lesions.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang Province, China.
- Department of Radiology, Putuo People's Hospital, School of Medicine, Tongji University, Shanghai, China.
- Department of Pathology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang Province, China.
- Zhejiang Academy of Traditional Chinese Medicine, Hangzhou, Zhejiang Province, China.
Abstract
Detection of clinically significant prostate cancer (csPCa) within PI-RADS category 3 lesions remains a major diagnostic challenge. To develop and validate a biparametric MRI (bpMRI)-based habitat analysis model integrating deep learning features for predicting csPCa in PI-RADS 3 lesions using dual-center data. Retrospective. This study included 551 patients with MRI-identified PI-RADS category 3 lesions and histopathological confirmation. A total of 439 patients from Center 1 were randomly assigned to a training set (n = 328) and an internal validation (in-vad) set (n = 111), while an external validation (ex-vad) set (n = 112) was obtained from Center 2. 3 T/1.5 T. T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) sequences. Lesions were manually segmented on preoperative T2WI and DWI, and tumor subregions were determined using k-means clustering. Deep learning features were obtained from each habitat subregion, and habitat-based models were built based on selected features. A habitat whole-tumor (Habitat W) model was subsequently derived by integrating all subregions. Recursive feature elimination (RFE) was applied to select the optimal predictors from the clinical and habitat-derived features; the clinical model was constructed using the selected clinical features, while the combined model incorporated all selected features. Student's t-test, Mann-Whitney U tests, Chi-squared tests, LASSO, areas under the curve (AUC), decision curve analysis (DCA), calibration curves, RFE, SHapley Additive exPlanations (SHAP). Statistical significance was defined as p-value < 0.05. In the training, in-vad and ex-vad sets, the clinical model demonstrated AUC values of 0.893, 0.844, and 0.837, respectively. The habitat models (habitat 1, 2,3 and -W) achieved AUCs ranging from 0.857 to 0.952. The combined model yielded AUCs of 0.959, 0.963, and 0.949, respectively. The bpMRI-based deep learning Habitat W and combined model enables accurate assessment of csPCa in PI-RADS 3 lesions. 3.