Diffusion Kurtosis Imaging in Diagnosing Parkinson's Disease: A Preliminary Comparison Study Between Kurtosis Metric and Radiomic Features.

Abstract

Rationale and objectives: Parkinson's disease (PD) shows small structural changes in nigrostriatal pathways, which can be sensitively captured through diffusion kurtosis imaging (DKI). However, the value of DKI and its radiomic features in the classification performance of PD still need confirmation. This study aimed to compare the diagnostic efficiency of DKI-derived kurtosis metric and its radiomic features with different machine learning models for PD classification.

Materials and methods: 75 people with PD and 80 healthy individuals had their brains scanned using DKI. These images were pre-processed and the standard atlas were non-linearly registered to them. With the labels in atlas, different brain regions in nigrostriatal pathways, including the caudate nucleus, putamen, pallidum, thalamus, and substantia nigra, were chosen as the region of interests (ROIs) to warped to the native space to measure the mean kurtosis (MK). Additionally, new radiomic features were developed for comparison. To handle the large amount of data, a statistical method called Z-score normalization and another method called LASSO regression were used to simplify the information. From this, a few most important features were chosen, and a combined score called Radscore was calculated using LASSO regression. For the comprehensive analyses, three different conventional machine learning models were then created: logistic regression (LR), support vector machine (SVM), and random forest (RF). To ensure the models were accurate, a process called 10-fold cross-validation was used, where the data were split into 10 parts for training and testing.

Results: Using MK alone, the models achieved good results in correctly identifying PD in the validation set, with LR at 0.90, RF at 0.93, and SVM at 0.90. When the radiomic features were added, the models performed even better, with LR at 0.92, RF at 0.95, and SVM at 0.91. Additionally, a nomogram combining all the information was created to predict the likelihood of someone having PD, which had an AUC of 0.91.

Conclusion: These findings suggest that the combination of DKI measurements and radiomic features can effectively diagnose PD by providing more detailed information about the brain's condition and the processes involved in the disease.