It is reported that radiomic features extracted from quantitative susceptibility mapping (QSM) had promising clinical value for the diagnosis of Parkinson’s disease (PD). We aimed to explore the usefulness of radiomics features based on magnitude images to distinguish PD from non-PD controls.
We retrospectively recruited PD patients and controls who underwent brain 3.0T MR including susceptibility-weighted imaging (SWI). A total of 396 radiomics features were extracted from the SN of 95 PD patients and 95 non-PD controls based on SWI. Intra-/inter-observer correlation coefficients (ICCs) were applied to measure the observer agreement for the radiomic feature extraction. Then the patients were randomly grouped into training and validation sets in a ratio of 7:3. In the training set, the maximum correlation minimum redundancy algorithm (mRMR) and the least absolute shrinkage and selection operator (LASSO) were conducted to filter and choose the optimized subset of features, and a radiomics signature was constructed. Moreover, radiomics signatures were constructed by different machine learning models. Area under the ROC curves (AUCs) were applied to evaluate the predictive performance of the models. Then correlation analysis was performed to evaluate the correlation between the optimized features and clinical factors.
The intro-observer CC ranged from 0.82 to 1.0, and the inter-observer CC ranged from 0.77 to 0.99. The LASSO logistic regression model showed good prediction efficacy in the training set [AUC = 0.82, 95% confidence interval (CI, 0.74–0.88)] and the validation set [AUC = 0.81, 95% CI (0.68–0.91)]. One radiomic feature showed a moderate negative correlation with Hoehn-Yahr stage (
Radiomic predictive features based on SWI magnitude images could reflect the Hoehn-Yahr stage of PD to some extent.