Abstracts

Rationale: Temporal lobe epilepsy (TLE) is associated with microstructural abnormalities extending beyond the medial temporal region. Our group recently demonstrated that Diffusional Kurtosis Imaging (DKI) Magnetic Resonance Imaging (MRI) is very sensitive to the microstructural abnormalities present in TLE. DKI is assessed by measuring non-Gaussian water diffusion properties, which are likely altered by pathological processes such as cell loss, glial infiltration and inflammation. Compared to conventional diffusion MRI metrics such as mean diffusivity (dMean) and fractional anisotropy (fa), mean kurtosis (kMean) is not only more intensely abnormal in TLE, but it also discloses abnormalities in areas typically not identified by fa or dMean. It is unknown, however, whether DKI abnormalities are consistently present in all patients with TLE. In this study, we performed a machine learning based classification of patients with left TLE versus controls using a data derived from kMean, dMean, and fa. We hypothesized that classification of TLE based on Kmean would be more accurate than classification from other modalities. We also predicted that combining all modalities would lead to an overall improvement in classification.Methods: We conducted a voxel-based study using diffusion MR imaging on a cohort of 36 healthy controls and 32 L TLE patients from a previously published report (Bonilha et al AJNR 2015 Apr, PMID: 25500311). Three modalities were used to compare patients versus controls: DTI-derived fa, DTI-derived (dMean), and DKI-derived (kMean). An SVM-based classifier was trained to classify patients with TLE vs. controls using a k-folds algorithm with n over 5 folds. Only data from white matter regions were employed. SVM was performed for each modality alone and for all modalities in combination. After each prediction, we identified brain regions that were most contributory to the classification.Results: By combining all the modalities, the SVM-based classifier was able to predict if a patient had epilepsy with an F1 score equal to 0.823. Sensitivity was 0.781 and specificity was .917 for this combined result. For each individual modality, the classifier F1 score was .699, .656, and .684 for kMean, dMean, and fa, respectively. The most contributory voxels for each of the modalities are displayed in Figure 1. The highest scoring modality, kMean, had a sensitivity of .688 and a specificity of .778.Conclusions: KMean demonstrated a higher accuracy in classifying patients with TLE compared with controls, suggesting that non-Gaussian microstructural abnormalities are better predictors of neuropathological changes associated with epilepsy. Nonetheless, the classification accuracy of each modality was only moderate and a stronger predictive value was obtained by combining data from modalities. This result suggests that there is likely an interplay between each biophysical measurement seen in epileptic change. Machine-based learning combined with voxel-based diffusion MRI shows promising potential for identifying structural abnormalities uniquely present in epilepsy.