Talai Aron S, Sedlacik Jan, Boelmans Kai, Forkert Nils D
Department of Radiology, University of Calgary, Calgary, AB, Canada.
Department of Diagnostic and Interventional Neuroradiology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany.
Front Neurol. 2021 Apr 14;12:648548. doi: 10.3389/fneur.2021.648548. eCollection 2021.
Patients with Parkinson's disease (PD) and progressive supranuclear palsy Richardson's syndrome (PSP-RS) often show overlapping clinical features, leading to misdiagnoses. The objective of this study was to investigate the feasibility and utility of using multi-modal MRI datasets for an automatic differentiation of PD patients, PSP-RS patients, and healthy control (HC) subjects. T1-weighted, T2-weighted, and diffusion-tensor (DTI) MRI datasets from 45 PD patients, 20 PSP-RS patients, and 38 HC subjects were available for this study. Using an atlas-based approach, regional values of brain morphology (T1-weighted), brain iron metabolism (T2-weighted), and microstructural integrity (DTI) were measured and employed for feature selection and subsequent classification using combinations of various established machine learning methods. The optimal machine learning model using regional morphology features only achieved a classification accuracy of 65% (67/103 correct classifications) differentiating PD patients, PSP-RS patients, and HC subjects. The optimal machine learning model using only quantitative T2 values performed slightly better and achieved an accuracy of 75.7% (78/103). The optimal classifier using DTI features alone performed considerably better with 95.1% accuracy (98/103). The optimal multi-modal classifier using all features also achieved an accuracy of 95.1% but required more features and achieved a slightly lower F1-score compared to the optimal model using DTI features alone. Machine learning models using multi-modal MRI perform significantly better than uni-modal machine learning models using morphological parameters based on T1-weighted MRI datasets alone or brain iron metabolism markers based on T2-weighted MRI datasets alone. However, machine learnig models using regional brain microstructural integrity metrics computed from DTI datasets perform similar to the optimal multi-modal machine learning model. Thus, given the results from this study cohort, it appears that morphology and brain iron metabolism markers may not provide additional value for classification compared to using DTI metrics alone.
帕金森病（PD）患者和进行性核上性麻痹理查森综合征（PSP-RS）患者常常表现出重叠的临床特征，从而导致误诊。本研究的目的是探讨使用多模态MRI数据集自动区分PD患者、PSP-RS患者和健康对照（HC）受试者的可行性和实用性。本研究可获取来自45例PD患者、20例PSP-RS患者和38例HC受试者的T1加权、T2加权和扩散张量（DTI）MRI数据集。使用基于图谱的方法，测量脑形态（T1加权）、脑铁代谢（T2加权）和微观结构完整性（DTI）的区域值，并使用各种既定机器学习方法的组合进行特征选择和后续分类。仅使用区域形态特征的最佳机器学习模型在区分PD患者、PSP-RS患者和HC受试者时的分类准确率为65%（67/103个正确分类）。仅使用定量T2值的最佳机器学习模型表现稍好，准确率为75.7%（78/103）。仅使用DTI特征的最佳分类器表现相当出色，准确率为95.1%（98/103）。使用所有特征的最佳多模态分类器准确率也达到了95.1%，但需要更多特征，并且与仅使用DTI特征的最佳模型相比，F1分数略低。使用多模态MRI的机器学习模型比仅使用基于T1加权MRI数据集的形态学参数或仅使用基于T2加权MRI数据集的脑铁代谢标志物的单模态机器学习模型表现显著更好。然而，使用从DTI数据集中计算出的区域脑微观结构完整性指标的机器学习模型与最佳多模态机器学习模型表现相似。因此，根据本研究队列的结果，与单独使用DTI指标相比，形态学和脑铁代谢标志物在分类方面可能无法提供额外价值。
