Jiang Minghan, Miao Zeyang, Xu Run, Guo Mengyao, Li Xuefeng, Li Guanwu, Luo Peng, Hu Su
Department of Radiology, Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shanghai University of Traditional Chinese Medicine, Shanghai, China.
Department of Radiology, The First Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China.
Front Oncol. 2025 Aug 6;15:1625158. doi: 10.3389/fonc.2025.1625158. eCollection 2025.
This study aimed to develop MRI-based radiomics machine learning models for predicting adverse pathological prognostic features in prostate cancer and to explore the feasibility of integrating radiomics with clinical characteristics to improve preoperative risk stratification, addressing the limitations of conventional clinical models.
A retrospective cohort of 137 prostate cancer patients between January 2021 and April 2023 with preoperative MRI and postoperative pathology data was divided into adverse-feature-positive (n=85) and negative (n=52) groups. Regions of interest (ROIs) were delineated on ADC and T2WI sequences, and 31 radiomics features were extracted using PyRadiomics. LASSO regression selected optimal features, followed by model construction via five algorithms (logistic regression, decision tree, random forest, SVM, AdaBoost). Clinical models incorporated three variables: biopsy Gleason grade, total PSA, and prostate volume. The best-performing radiomics model was combined with clinical features to build a hybrid model. Model performance was evaluated by AUC, sensitivity, specificity, accuracy, calibration curves, and decision curve analysis (DCA).
Patients were randomly split into training (n=95) and validation (n=42) cohorts. The random forest model using ADC-T2WI combined features achieved the highest AUC (0.832; 95% CI: 0.706-0.958) in the validation set, outperforming the clinical model (AUC=0.772). The hybrid model demonstrated superior performance (AUC=0.909; 95% CI: 0.822-0.995), with sensitivity=0.813, specificity=0.885, and accuracy=0.857. Calibration and DCA confirmed its robust clinical utility (<0.01 . single models).
The biparametric MRI radiomics-random forest model effectively predicts adverse pathological features in prostate cancer. Integration with clinical characteristics further enhances predictive accuracy, offering a non-invasive tool for preoperative risk stratification and personalized treatment planning.
本研究旨在开发基于MRI的放射组学机器学习模型，以预测前列腺癌的不良病理预后特征，并探索将放射组学与临床特征相结合以改善术前风险分层的可行性，解决传统临床模型的局限性。
回顾性队列研究纳入了2021年1月至2023年4月期间137例有术前MRI和术后病理数据的前列腺癌患者，分为不良特征阳性组（n = 85）和阴性组（n = 52）。在ADC和T2WI序列上勾画感兴趣区（ROI），使用PyRadiomics提取31个放射组学特征。LASSO回归选择最佳特征，随后通过五种算法（逻辑回归、决策树、随机森林、支持向量机、AdaBoost）构建模型。临床模型纳入三个变量：活检Gleason分级、总PSA和前列腺体积。将表现最佳的放射组学模型与临床特征相结合构建混合模型。通过AUC、敏感性、特异性、准确性、校准曲线和决策曲线分析（DCA）评估模型性能。
患者被随机分为训练组（n = 95）和验证组（n = 42）。在验证集中，使用ADC-T2WI联合特征的随机森林模型获得了最高的AUC（0.832；95%CI：0.706 - 0.958），优于临床模型（AUC = 0.772）。混合模型表现出卓越的性能（AUC = 0.909；95%CI：0.822 - 0.995），敏感性 = 0.813，特异性 = 0.885，准确性 = 0.857。校准和DCA证实了其强大的临床实用性（<0.01，优于单一模型）。
双参数MRI放射组学 - 随机森林模型能有效预测前列腺癌的不良病理特征。与临床特征相结合进一步提高了预测准确性，为术前风险分层和个性化治疗规划提供了一种非侵入性工具。
