This project will test the accuracy of a novel diffusion-weighted magnetic resonance imaging (DWMRI) approach using a deep convolutional neural network (DCNN) to predict an optimal resection margin for pediatric epilepsy surgery objectively. Its primary goal is to minimize surgical risk probability (i.e., functional deficit) and maximize surgical benefit probability (i.e., seizure freedom) by precisely localizing eloquent white matter pathways in children and adolescents with drug-resistant focal epilepsy. This new imaging approach, which will acquire a DWMRI scan before pediatric epilepsy surgery in about 10 minutes without contrast administration (and also without sedation even in young children), can be readily applied to improve preoperative benefit-risk evaluation for pediatric epilepsy surgery in the future. The investigators will also study how the advanced DWMRI-DCNN connectome approach can detect complex signs of brain neuronal reorganization that help improve neurological and cognitive outcomes following pediatric epilepsy surgery. This new imaging approach could benefit targeted interventions in the future to minimize neurocognitive deficits in affected children. All enrolled subjects will undergo advanced brain MRI and neurocognitive evaluation to achieve these goals. The findings of this project will not guide any clinical decision-making or clinical intervention until the studied approach is thoroughly validated.

This project will combine advanced brain MRI with detailed neuro-psychology evaluation, performed in children and adolescents affected by drug-resistant focal epilepsy, to address two main aims, each of them with the following research hypotheses:

AIM 1. To determine the accuracy of deep learning tractography-based benefit-risk analysis compared to a standard electrical stimulation mapping (ESM) which is the current clinical standard for detecting eloquent cortical regions before epilepsy surgery.

Hypothesis 1.1 In healthy controls, DCNN-based tract classification will localize eloquent cortices, which are significantly overlapped at both single shell DWI acquisition and generalized Q-sampling imaging, suggesting that the accuracy of this approach may not be significantly affected by the acquisition protocol.

Hypothesis 1.2 DCNN-based tract classification will achieve at least 93% accuracy for prospective detection of ESM-defined eloquent cortices, including patients with a high likelihood of functional reorganization.

Hypothesis 1.3 Preservation of DCNN-classified eloquent white matter pathways during surgery will predict the avoidance of postoperative deficits as accurately as the preservation of ESM-defined eloquent cortex.

Hypothesis 1.4 Preservation of surgical margins optimized by Kalman filter on retrospective data will achieve seizure control and avoidance of postoperative deficits in a prospective surgical patient cohort.

In this aim, the investigators will test the accuracy of a recently developed deep learning-based benefit-risk model, called deep convolutional neural network (DCNN) tract classification combined with Kalman filter analysis, for non-invasive detection of eloquent white matter pathways and optimization of surgical margin (i.e., the distance between epileptogenic area and eloquent area), resulting in seizure freedom and avoidance of functional deficits. Failure to identify eloquent areas in the proposed resection region can have potentially lifelong consequences, and overestimation or incorrect localization of the extent of the eloquent regions may lead to incomplete resection of the epileptogenic zone. Without optimizing the benefit-risk ratio, the minimum acceptable margin is highly variable across different settings, ranging from 0 to 2 cm across epilepsy surgery centers. The investigators will study whether the proposed benefit-risk model can standardize (or customize) epilepsy surgery of individual patients by accurately optimizing the margins of the eloquent white matter pathways to be preserved, which is ultimately essential to balance the benefit of seizure freedom with the risk of functional deficit. This proposed new imaging approach could change clinical practice for pediatric epilepsy surgery and is widely applicable for other types of neurosurgical procedures such as tumor resection.

AIM 2. To determine the accuracy of deep learning-based connectome analysis for prediction of long-term neurocognitive improvement following epilepsy surgery.

Hypothesis 2.1 Connectivity efficiencies preserved in specific modular networks of preoperative DCNN-based connectome, found to be associated with postoperative functional improvement on retrospective data, will accurately predict long-term functional improvement in a prospective patient cohort.

Hypothesis 2.2 Longer epilepsy duration will be significantly associated with more decreased efficiency in full-scale IQ modular network of preoperative DCNN-based connectome, thus suggesting that earlier surgery will yield better long-term full-scale IQ improvement.

Hypothesis 2.3 Patients with ipsilateral resections, who show signs of postoperative "crowding" (i.e., verbal IQ improvement at the expense of non-verbal function), will show decreased efficiency in non-verbal and increased efficiency in verbal IQ network of DCNN-based connectome in the contralateral hemisphere.

In this aim, the investigators will test if an advanced DWMRI approach integrating DCNN and connectome helps decide timely surgery by providing 1) preoperative imaging markers underlying high likelihood of postoperative neurocognitive improvements and 2) mechanistic insight in structural brain reorganization associated with postoperative verbal IQ improvement. A series of preoperative imaging markers called "local efficiency" that quantifies how efficiently neural connection is shared by neighboring regions will be evaluated at the levels of specific modular networks. We expect that these markers can identify long-term and specific neurocognitive consequences (and potential predictors of these) associated with surgical intervention and their neural correlates for specific neurocognitive functions. In addition, neuronal remodeling associated with a functional crowding effect, studied with DWMRI connectome improved by the DCNN tract classification, will provide a new mechanistic insight in compensatory processes for verbal IQ function in children and adolescents who undergo resective surgery to treat drug-resistant focal epilepsy.

Inclusion Criteria:

1. Subjects with drug-resistant focal epilepsy

   1. Age 3-19 years. 2. Planned two-stage epilepsy surgery with subdural electrodes.

2. Healthy control subjects 1. Age 5-19 years. 2. No cognitive, motor, and/or language impairment or clinical elevations on a measure of behavioral problems. 3. Brain MRI interpreted as normal.

Exclusion Criteria:

For all subjects:

1. History of prematurity or perinatal hypoxic-ischemic event. 2. Hemiplegia on preoperative neurological examination by pediatric neurologists. 3. Dysmorphic features suggestive of a clinical syndrome. 4. Diagnosis of any pervasive developmental or psychiatric condition which clearly predates the onset of seizures, including autism spectrum disorder, tic disorders, obsessive-compulsive disorder. 5. MRI abnormalities showing massive brain malformation and other extensive lesions that likely destroyed the contralateral tracts and severely affected i) spatial normalization accuracy in advanced normalization tools (ANTs), mutual information (MI) between native T1- MRI of Geodesic SyN transform and template T1-MRI < mean-3*standard deviation of MI in the healthy control group and ii) parcellation accuracy in surface-matching-based deformable registration, target registration error (TRE) of fine tetrahedra mesh between native T1- MRI brain surface and template T1-MRI brain surface > mean-3*standard deviation of TRE in the healthy control group. 6. History of claustrophobia. 7. Unsuccessful MRI showing head motion > 2 mm in DWMRI (i.e., voxel size of DWMRI) which is evaluated by NIH TORTOISE DWMRI motion artifact correction package. 8. Subject who cannot speak English.