A multiple-tissue-specific magnetic resonance imaging model for diagnosing Parkinson's disease: A brain radiomics study

Abstract

Brain radiomics can reflect the characteristics of brain pathophysiology. However, the value of T1-weighted images, quantitative susceptibility mapping, and R2∗ mapping in the diagnosis of Parkinson's disease (PD) was underestimated in previous studies. In this prospective study to establish a model for PD diagnosis based on brain imaging information, we collected high-resolution T1-weighted images, R2∗ mapping, and quantitative susceptibility imaging data from 171 patients with PD and 179 healthy controls recruited from August 2014 to August 2019. According to the inclusion time, 123 PD patients and 121 healthy controls were assigned to train the diagnostic model, while the remaining 106 subjects were assigned to the external validation dataset. We extracted 1408 radiomics features, and then used data-driven feature selection to identify informative features that were significant for discriminating patients with PD from normal controls on the training dataset. The informative features so identified were then used to construct a diagnostic model for PD. The constructed model contained 36 informative radiomics features, mainly representing abnormal subcortical iron distribution (especially in the substantia nigra), structural disorganization (e.g., in the inferior temporal, paracentral, precuneus, insula, and precentral gyri), and texture misalignment in the subcortical nuclei (e.g., caudate, globus pallidus, and thalamus). The predictive accuracy of the established model was 81.1 ± 8.0% in the training dataset. On the external validation dataset, the established model showed predictive accuracy of 78.5 ± 2.1%. In the tests of identifying early and drug-naïve PD patients from healthy controls, the accuracies of the model constructed on the same 36 informative features were 80.3 ± 7.1% and 79.1 ± 6.5%, respectively, while the accuracies were 80.4 ± 6.3% and 82.9 ± 5.8% for diagnosing middle-to-late PD and those receiving drug management, respectively. The accuracies for predicting tremor-dominant and non-tremor-dominant PD were 79.8 ± 6.9% and 79.1 ± 6.5%, respectively. In conclusion, the multiple-tissue-specific brain radiomics model constructed from magnetic resonance imaging has the ability to discriminate PD and exhibits the advantages for improving PD diagnosis.