An Interpretable Clinical-Radiomics-Deep Learning Model Based on Magnetic Resonance Imaging for Predicting Postoperative Gleason Grading in Prostate Cancer: A Dual-Center Study

Fuyu Guo¹Shiwei Sun^2,3Xiaoqian Deng¹Yue Wang⁴Wei Yao⁵Peng Yue⁶Yangang Zhang^1,7*Yanbin Liu^3*Yingzhong Yang^3*

• ¹Third Hospital of Shanxi Medical College, Taiyuan, China
• ²Department of Urology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China
• ³Department of Urology, Taiyuan Central Hospital of Shanxi Medical University, Taiyuan, 030000, China., Taiyuan, Shanxi Province, China
• ⁴Department of Urology, Xinzhou People's Hospita, Xinzhou, Shanxi Province, China
• ⁵Department of Urology, Datong Fifth People's Hospital, Datong, Shanxi Province, China
• ⁶Department of Urology, Handan First Hospital, Handan, Hebei Province, China
• ⁷Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Tongji Shanxi Hospital, Third Hospital of Shanxi Medical University, Taiyuan, Shanxi Province, China

Objective: To develop and test an interpretable machine learning model that combines clinical data, radiomics, and deep learning features using different regions of interest (ROI) from magnetic resonance imaging (MRI) to predict postoperative Gleason grading in prostate cancer (PCa). Methods: A retrospective analysis was conducted on 96 PCa patients from the Third Hospital of Shanxi Medical University (training set) and 33 patients from Taiyuan Central Hospital (testing set) treated between August 2014 and July 2022. Clinical data, including prostate-specific antigen and MRI data, were collected. Tumor and whole-prostate ROIs were delineated on T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient sequences. Following image preprocessing, traditional radiomics and deep learning features were extracted and combined with clinical features. Various machine learning models were constructed using feature selection methods such as LASSO regression. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves (CALC), decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP) analysis. Results: All combined models performed well in the test set (AUC ≥0.75), with the LightGBM model achieving the highest accuracy (0.848). SHAP analysis effectively illustrated the contribution of each feature. The CALC demonstrated good agreement between predicted probabilities and actual outcomes, and DCA further indicated that the models provided significant net benefits for clinical decision-making across various risk thresholds. Conclusion: This study developed and validated interpretable MRI-based machine learning models that combine clinical data with radiomics and deep learning features from different regions of interest, demonstrating good performance in predicting postoperative Gleason grading in PCa.