磁共振成像2025，Vol.16Issue(12)：22-29,8.DOI:10.12015/issn.1674-8034.2025.12.004
基于可解释性机器学习与T1成像核团特征的帕金森患者诊断和分期模型研究
Research on diagnostic and staging models for Parkinson's disease patients using T1 images based on interpretable machine learning
摘要
Abstract
Objective:To construct a three-classification model based on T1-weighted imaging(T1WI)radiomics for the diagnosis and early,middle and late stage classification of Parkinson's disease(PD),and to explore the diagnostic value of different nuclear features.Materials and Methods:A prospective analysis was conducted on the T1WI data of 146 patients,including 86 cases from the Second Affiliated Hospital of Xinjiang Medical University(development cohort),with 26 cases in the normal group,35 cases in the early PD group,and 25 cases in the middle and late PD group;and 60 cases from Affiliated First Hospital of Xinjiang Medical University(external validation cohort),with 18 cases in the normal group,22 cases in the early PD group,and 20 cases in the middle and late PD group.Six nuclei,namely the caudate nucleus(CN),putamen(PUT),globus pallidus(GP),red nucleus(RN),substantia nigra(SN),and nucleus accumbens(NAC),were segmented from all data,and 1688 radiomics features(including first-order statistics,shape,texture,and filter features)were extracted.Logistic regression(LR)algorithm was used to construct 6 single-nucleus models and 1 combined model.Key features were selected through variance threshold method,univariate selection method,and least absolute shrinkage and selection operator(LASSO)algorithm.Model performance and interpretability were analyzed using receiver operator characteristic(ROC)curve,confusion matrix,and SHapley Additive exPlanations(SHAP)value.Results:The macro area under the curve(AUC)of the combined model in the training set,internal validation set,and external validation set were 0.93(95%CI:0.87 to 1.00),0.88(95%CI:0.69 to 1.00),and 0.84(95%CI:0.75 to 0.92),respectively,which were significantly better than most single-nucleus models.Among the 21 key features selected,9 were related to GP(e.g.,wavelet-HLL_glcm_Correlation_GP),and the absolute value of the coefficient of the lbp-3D-k_firstorder_Minimum_RN feature of RN was the largest(0.12).SHAP analysis showed that the classification of the normal group relied on the texture symmetry of GP and SN,early PD focused on the local structural changes of PUT and GP,and middle and late PD was characterized by signal abnormalities of RN and NAc.Conclusions:The multi-nucleus combined model based on T1WI has high performance in the diagnosis and staging of PD.Nucleus-specific features can reflect the spatial heterogeneity of the pathological process of PD,with GP and SN playing a core role in diagnosis,providing an imaging basis for precise clinical staging.关键词
帕金森病/磁共振成像/T1加权成像/可解释性机器学习/核团特征/诊断分期Key words
Parkinson's disease/magnetic resonance imaging/T1-weighted imaging/explainable machine learning/nucleus features/diagnostic staging分类
医药卫生引用本文复制引用
HAO Lu,ZHU Minghui,ZHU Yutong,KALIBUNUER·Mahemuti,WANG Yunling,HANJIAERBIEKE·Kukun,GUAN Yangtai..基于可解释性机器学习与T1成像核团特征的帕金森患者诊断和分期模型研究[J].磁共振成像,2025,16(12):22-29,8.基金项目
Natural Science Foundation Project of Xinjiang Uygur Autonomous Region(No.2024D01C147).新疆维吾尔自治区自然科学基金项目(编号:2024D01C147) （No.2024D01C147）
