Convolutional neural network optimizes the application of diffusion kurtosis imaging in Parkinson's disease.

Objectives: The literature regarding the use of diffusion-tensor imaging-derived metrics in the evaluation of Parkinson's disease (PD) is controversial. This study attempted to assess the feasibility of a deep-learning-based method for detecting alterations in diffusion kurtosis measurements associated with PD.

Methods: A total of 68 patients with PD and 77 healthy controls were scanned using scanner-A (3 T Skyra) (DATASET-1). Meanwhile, an additional five healthy volunteers were scanned with both scanner-A and an additional scanner-B (3 T Prisma) (DATASET-2). Diffusion kurtosis imaging (DKI) of DATASET-2 had an extra b shell compared to DATASET-1. In addition, a 3D-convolutional neural network (CNN) was trained from DATASET-2 to harmonize the quality of scalar measures of scanner-A to a similar level as scanner-B. Whole-brain unpaired t test and Tract-Based Spatial Statistics (TBSS) were performed to validate the differences between the PD and control groups using the model-fitting method and CNN-based method, respectively. We further clarified the correlation between clinical assessments and DKI results.

Results: An increase in mean diffusivity (MD) was found in the left substantia nigra (SN) in the PD group. In the right SN, fractional anisotropy (FA) and mean kurtosis (MK) values were negatively correlated with Hoehn and Yahr (H&Y) scales. In the putamen (Put), FA values were positively correlated with the H&Y scales. It is worth noting that these findings were only observed with the deep learning method. There was neither a group difference nor a correlation with clinical assessments in the SN or striatum exceeding the significance level using the conventional model-fitting method.

Conclusions: The CNN-based method improves the robustness of DKI and can help to explore PD-associated imaging features.