Abstracts

Rationale: Detection of cerebral abnormalities is an important step for the presurgical evaluation of patients with medically refractory focal epilepsy. A machine-learning-enabled, high-dimensional method of detection employing multimodal MRI data may be able to help detect and localize such abnormalities. Methods: We recruited 62 controls, 33 patients with discrete MRI lesions (MRI positive), and 23 patients with no overt MRI lesions (MRI negative). We developed a high-dimensional, machine-learning-enabled method of cerebral abnormality detection based on regional and voxel-based brain MRI features, and trained it with multimodal MRI, and MRI-derived data (3D-T1, T2, 3D-FLAIR, DTI radial diffusivity, DTI fractional anisotropy, DTI axial diffusivity, NODDI intercellular volume fraction, susceptibility-weighted angiography, DESPOT T1, DESPOT PD1,2, and cortical thickness3) from control subjects and patients with discrete lesions. Evaluation was performed on discrete lesions and control data, and the MRI negative cases. Results: In discrete lesion cases, the mean dice score coefficient measuring the degree of spatial overlap between manually drawn lesion masks and the brain abnormality masks was 0.628 (standard deviation: 0.195), as measured on held-out test data (Figure 1). There were no lesional areas detected in control cases. In the MRI negative cohort, there was qualitative agreement in 16 out of 23 cases between detected lesional areas and the results of clinical investigations such as scalp EEG, neuropsychology, and FDG PET imaging (Figure 2). In the remaining 7 MRI negative cases, there was no clear concordance/disconcordance. Conclusions: This approach appears to be promising for detecting cerebral abnormalities in both MRI positive and negative pharmacoresistant focal epilepsy. Further development and evaluation, including comparison with the results of forthcoming stereo EEG investigations are planned. Funding: This work was supported by the Medical Research Council (MRC), Sobell Foundation, and the National Institute for Health Research (NIHR) University College London NHS Foundation Trust (UCLH) Biomedical Research Centre (BRC).