Abstracts

Rationale:
Many investigators have attempted to identify non-invasive imaging markers allowing more accurate quantification of epileptogenicity based on anatomical and functional MRI1,2. The present study investigated whether a state-of-the-art multi-scale deep learning network would optimize the localization of seizure onset zone (SOZ) by effectively predicting the regional likelihood of epileptogenicity from clinically acquired multi-modal MRI of children with focal epilepsy.
Method:
We studied 42 children with drug-resistant focal epilepsy (mean age: 10.3 years, cortical dysplasia/tumor/non-specific histology: 15/9/18) in whom ictal ECoG successfully localized SOZ sites and resective surgery resulted in the ILAE Class 1 outcome. A cortical parcellation was applied to define 1002 cortical regions, Ri=1-1002, in the brain. For each Ri, the surface laminar analysis presented in our previous study3 was applied to sample the degree of atypical change in 1) gray matter morphology using 299 multimodal MRI values (scaled from T1-weighted, T2-weighted, FLAIR, DWI) and 2) white matter connectivity using 1002 DWI connectome edge strengths. The sampled features served as the input vector, xi, to predict two classes, C1: SOZ and C2: non-SOZ (non-epileptic zone), using a variant of a multi-scale residual neural network4 (ms-ResNet, Fig. 1). A 5-fold cross-validation using xi of the epileptogenic hemispheres was employed to train and test all network layers in an end-to-end fashion. A prediction probability value of a given xi belonging to C1 was estimated at the softmax layer to define seizure onset likelihood, μi.
Results:
The training set could provide successful convergence without artificial data augmentation (Fig. 2A). In the test set, the proposed ms-ResNet achieved an accuracy/sensitivity/specificity of 0.96/0.62/0.98 for correct classification (Fig. 2B), yielding a very large effect size for mi, Cohen’s d = 3.38, between SOZ and non-SOZ electrodes (Fig. 2C).
Conclusion:
While considering high-dimensional abnormalities relevant in both gray matter morphology and white matter connectivity, the proposed ms-ResNet could effectively learn specific signatures of multi-modal MRI in epileptogenic foci that have different types of histopathology. Our preliminary data showed that deep learning-based seizure onset likelihood of multi-modal MRI features could be highly specific to predict non-epileptic sites, suggesting a new non-invasive marker that may guide the placement of intracranial electrodes during clinical practice. Reference •Hong SJ, et al. Automated detection of cortical dysplasia type II in MRI-negative epilepsy. Neurology. 2014; 83(1):48-55. •Nedic S et al., Using network dynamic fMRI for detection of epileptogenic foci. BMC Neurol. 2015; 15:262. •Govindan RM, et al., Surface-based laminar analysis of diffusion abnormalities in cortical and white matter layers in neocortical epilepsy. Epilepsia. 2013; 54:667-77. •Wang F et al., CSI-Net: Unified human body characterization and pose recognition, arXiv:1810.03064.
Funding:
:NINDS R01NS089659 (to J.J.) and R01NS064033 (to E.A.),