Assessing microstructural brain differences in epileptic patients with vagus nerve stimulation via diffusion MRI, tractography and machine learning

Objective: Vagus Nerve Stimulation (VNS) offers treatment for drug-resistant epilepsy, but the mechanism underlying its efficacy remains unclear. This study aimed to uncover microstructural features in VNS responders and non-responders using diffusion magnetic resonance imaging (dMRI), diffusion models and machine learning techniques. Methods: Data were collected from 19 patients with drug-resistant epilepsy. Diffusion tensor imaging and other multi-compartment models (NODDI, DIAMOND and Microstructure Fingerprinting) were used to compute microstructural metrics within distinct regions of interest (ROIs) outlined on tractography pathways estimated using FreeSurfer and MRtrix3. We calculated the weighted mean, standard deviation, skewness, and kurtosis of the metrics along the tracts for summarising their distribution. Univariate analysis has been performed through non-parametric statistical tests: Mann-Whitney U rank, Kruskal-Wallis and Barnard exact tests. Multivariate analysis has been performed by a sequential feature selector to select the best classifying set of microstructural metrics. The extraction of Radiomics features has been conducted to get more informative characteristics about the shape and voxel intensities of the selected regions, feature selection algorithms together with classification algorithms have allowed us to classify non-responder patients. Deep Learning methodologies have been applied to classify patients without the use of precomputed ROIs. A pre-trained 3D encoder was used to reduce the size of the volumes and classify the responsiveness. Results: Treatment in non-responders demonstrated a greater mean diffusivity (MD) in the thalamocortical connections, the fornix, and the anterior commissure (p < 0.01), as well as the feature selector, selected the fornix as best classifying features. Wavelet and local binary pattern features have been the most frequently selected by the Radiomics pipeline, reaching an accuracy above 0.9. Expected results have been found with deep learning approaches, overfitting has been observed due to the lack of a large dataset. Interpretation: Further studies are needed to fully understand the roles of the corpus callosum, anterior commissure and longitudinal fasciculus in mediating the effect of VNS. Our results emphasize the potential of microstructural connections, machine learning and deep learning to guide personalized VNS treatment adjustments and prediction of non-responders patients to VNS.