mpMRI-based clinic-radiomics-deep learning model integrating lesion and PPAT for predicting csPCa in PI-RADS category 3 lesions: a multicenter study.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, the First Affiliated Hospital of Anhui Medical University, Hefei, China.
- Department of Radiology, Shanghai Pudong New Area Gongli Hospital, Shanghai, China.
- Department of Radiology, Affiliated Suzhou Hospital, Anhui Medical University, Suzhou, China.
- Department of Radiology, The First Affiliated Hospital of Xi'an Medical Universiy, Xi'an, China.
- Department of Radiology, the First Affiliated Hospital of Anhui Medical University, Hefei, China. [email protected].
- Department of Radiology, The First Affiliated Hospital of Xi'an Medical Universiy, Xi'an, China. [email protected].
- Department of Radiology, the First Affiliated Hospital of Anhui Medical University, Hefei, China. [email protected].
Abstract
Prostate PI-RADS category 3 lesions are indeterminate, requiring accurate tools to predict benign/malignant status and avoid unnecessary biopsies. The purpose of the study is to develop a clinical-radiomics-deep learning fusion model integrating lesions and periprostatic adipose tissue (PPAT) based on multi-parametric magnetic resonance imaging (mpMRI) for improving the detection of clinically significant prostate cancer (csPCa) in PI-RADS category 3 lesions. A total of 386 patients diagnosed with PI-RADS category 3 lesions via MRI were enrolled in this study. The cohort consisted of a training set (n=208) and an internal validation set (n=89) from Center A, as well as an external test set (n=88) from Centers B and C. Based on preprocessed prostate T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) images, prostate lesions and PPAT were separately delineated as volumes of interest (VOIs). Radiomic and deep learning features were then extracted and screened using Pearson correlation analysis, least absolute shrinkage and selection operator (LASSO) and deep feature compression. Finally, features derived from lesions and PPAT were subjected to feature concatenation and combined with specific clinical parameters to construct the final multimodal fusion model. The performance of this model was evaluated in the internal validation cohort and external test cohort using receiver operating characteristic (ROC) curves. The fusion model integrating lesions and PPAT outperformed all single-modality models. Its AUC was 0.954 (95%CI 0.923-0.985) in the training set, 0.895 (95%CI 0.814-0.976) in the internal validation set, and 0.800 (95%CI 0.694-0.906) in the external test set. Accuracy, sensitivity, and specificity were 0.822, 0.850, and 0.814 (internal validation) and 0.818, 0.679, and 0.883 (external test), respectively. The clinical-radiomics-DL fusion model based on lesions and PPAT shows promising performance for identifying csPCa in PI-RADS category 3 lesions under multicenter retrospective validation. It should be regarded as investigational auxiliary decision-making-support tool rather than ready-for-routine-clinical-use tool.