Early diagnosis of Parkinson's Disease (PD) has recently attracted extensive focus for its importance. However, most of the existing research only uses single-modal(gray matter, white matter or cerebrospinal fluid). In this study, we proposed a methodological framework to distinguish early PD patients from normal controls. This approach involved data analysis from integrated levels of structure. For each matric, we computed the values of 116 regions of interest derived from a prior atlas, which use principal components analysis(PCA) to reduce the dimensions, then trained with relevance vector machine(RVM)-based classifier. The performance of this method was evaluated using leave-one-out cross-validation. Applying the approach to a real data set containing 19 PD patients and 27 normal controls led to a classification accuracy of 89.13% with a sensitivity of 78.95% and a specificity of 96.30%. The proposed method shows promising classification performance by combining information from different levels, and it has the potential to improve the early clinical diagnosis and treatment evaluation of PD.