Track: 3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Jonah Isen
Undergraduate Researcher
School of Computing, Queen's University
Focal epilepsy is a common neurological disorder in which many patients are refractory to medical treatment. Surgery may be an option if the focus can be identified. Thus computational advances in pre-surgical evaluation are important. This study performs a voxel based analysis on patients with refractory focal epilepsy using multimodal MRI data to identify the causative brain lesion.
Methods:
This study used MRI data from 52 individuals with refractory focal epilepsy, 44 of which had visible lesions in MRI scans (discrete group) and eight that had normal MRI scans (negative group). These subjects were analyzed against 62 healthy control subjects. For each subject, T1-weighted MRI, diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI) were used. Specific utilized modalities included T1-segmented grey matter concentration (GMC), DTI fractional anisotropy (FA), DTI mean diffusivity (MD), NODDI neurite density index (NDI), and NODDI orientation dispersion (ODI). The multimodal data was first pre-processed using MATLAB code that used subroutines from SPM12. This pre-processing registered the diffusion imaging modalities to a subject’s T1, and subsequently normalized the data to MNI space and smoothed the images with 8mm gaussian FWHM. Two-tailed t-tests were then performed on each modality comparing individual subjects with epilepsy to controls. Resulting statistical maps were thresholded at t-values of three and visually compared to ground truth masks to ascertain the validity of lesion detection. Manually drawn lesion masks were used as ground truth for discrete subjects, while seizure onset zone (SOZ) masks from stereoelectroencephalography were used as ground truth for negative subjects. A leave-one-out cross validation analysis was furthermore performed on control subjects to test for false positives in controls.
Results:
In discrete subjects, lesion detection from decreased NDI demonstrated the best results, detecting accurate lesions in 76% of patients. Conversely, GMC provided the worst detection with both decreased and increased GMC only locating lesions in just under 50% of patients. Similar results were seen from negative subjects, with NDI again outperforming other modalities with 63% lesion detection, and GMC providing the worst results with 38% lesion detection for decreased GMC and 25% from increased GMC. Full results can be seen in Table 1. Analysis of controls demonstrated minimal significant findings across all modalities with average false positive voxel counts below 300.
Conclusions:
This work demonstrated the capability of different imaging modalities to accurately locate brain lesions in both discrete and negative patients. In most cases, lesion detection with NODDI appeared most promising for detecting focal lesions and was superior to conventional DTI. It appeared helpful both for visible lesions and in patients who are MRI negative supporting prior work (Winston 2015). Future work will apply multiparametric voxel-based analysis of several modalities simultaneously to improve sensitivity and specificity.
Funding: Please list any funding that was received in support of this abstract.: We are grateful to the Epilepsy Society for supporting the Epilepsy Society MRI scanner. This research was supported by the National Institute for Health Research University College London Hospitals Biomedical Research Centre. Data acquisition was funded by the Medical Research Council (MR/M00841X/1).