Keywords: Structural MRI, Machine learning, Epilepsy, Focal cortical dysplasia
Number of Figures: 7
Number of words (approximate): TBD

Introduction

Drug-resistant focal epilepsy is a disabling disease, leading to increased morbidity and mortality as well as psychosocial and economic impairment \cite{ McCagh2009,Mohanraj2006,Theodore2006}. In appropriately selected patients, surgery leads to long-term seizure freedom in approximately 65% and 45% of patients with temporal and extratemporal lobe epilepsy, respectively \cite{Téllez-Zenteno2005}. Seizure outcomes are significantly improved in patients with identifiable epileptogenic lesions on MRI \cite{Téllez-Zenteno2010}. While some epileptogenic lesions easy to detect, others are subtle or equivocal; in 15-30% of patients, no appreciable lesions are identified \cite{Duncan2016}. Approximately 14% of all patients undergoing epilepsy surgery and half of non-lesional patients have focal cortical dysplasia (FCD) by pathology\cite{Blumcke2017} (Chapman 2005, Lee 2005, McGonigal 2007 or Bien 2009, Lerner 2009); 50-80% of FCD patients have normal-appearing MRIs (Besson 2008). Therefore, improved identification of subtle focal cortical dysplasias (FCDs) is of great clinical importance and may improve surgical outcomes.
There is therefore significant interest in post-processing methods that allow for automated FCD detection. This task is challenging, however, both because the lesions can be subtle, and therefore similar in appearance to normal cortex, and because FCD lesions are heterogeneous. Histologically, findings range from subtle abnormalities of cortical lamination in FCD type I to the presence of large dysmorphic neurons with or without balloon cells in FCD type II \citep{Blümcke2011}.  Characteristic radiographic findings include changes in cortical thickness, T1 and/or T2/FLAIR weighted imaging intensity (including the transmantle sign in FCD type IIb), gray-white junction (GWJ) blurring, and cortical folding changes  (see reviews in Kini, Adler 2017b); although well-described, the appearance both within and across lesions is variable, even within type II lesions  \cite{Hong2017}
Two main post-processing approaches have been taken to aid with visual detection of FCDs.  The first is creation of contrast maps that highlight areas with increased likelihood, or similarity, to FCDs.  The most commonly used of these approaches is the Morphometric Analysis Program (MAP)  \cite{Huppertz2005}(Wagner 2011), a voxel-based morphometric approach that highlights areas of  GWJ blurring and altered signal intensity due to cortical thickening or abnormal gyration. Other FCD-specific univariate feature maps have also been found to be of use (see review Bernasconi 2011). More recently, several groups have reported successful fully automated FCD detection approaches using multivariate surface-based morphometry and machine learning  \cite{Hong2014,Adler2017}, Ahmed 2015.  Such fully automated approaches typically require large lesional and normative training sets, limiting their generalizability.
In this work, we contribute to this literature by using normative modeling to better understand the ways in which variability both of FCDs and of normal cortex may contribute to decreased sensitivity and/or specificity in these post-processing approaches.  We use a generic feature set, as opposed to FCD-specific features, and create a normative model using healthy volunteer data to represent the local neighborhood surrounding each voxel across multiple MR contrasts. We found that while FCD lesions do appear anomalous in our representation, so do several "normal" brain regions. In fact, we find that FCDs share many features with some of these outlying regions, particularly paralimbic cortical areas.  Within our normative model, we are able to create similarity maps to any selected region of interest; FCD similarity maps provide a useful contrast to identify similarities to other brain regions, and also aid with FCD identification, particularly after local normalization to account for similarities to these similar-appearing cortical regions.  We demonstrate the utility of our FCD contrast maps, as well as an automated detection method, in our case series.

Materials and Methods

Study Participants

From our surgical epilepsy imaging database, we retrospectively identified 15 consecutive patients undergoing presurgical evaluation for drug-resistant focal epilepsy with 1) radiologically apparent (MRI+) or histologically proven (MRI+ or MRI negative, MRI-) FCDs and 2) 3T MRI structural imaging using our standard NIH epilepsy imaging protocol from 2014-2019.  Patients were excluded if they had other MR imaging protocols, low image quality on visual inspection. The control group consisted of 30 healthy volunteers (HVs) scanned using the same imaging protocol, with no previous history of neurologic, psychiatric or other significant medical illnesses that may affect the central nervous system. Data were collected at the Clinical Center of the National Institutes of Health (NIH; Bethesda, MD). All participants were enrolled in a research protocol approved by the Institutional Review Board; informed consent was obtained from all participants.

Lesion Labels

For MRI+ patients, lesions were traced in the volume using AFNI by an experienced neurologist using the T1wi, informed by the T2wi and FLAIR images when necessary. For MRI- patients, the postoperative T1wi was registered to the preoperative T1wi in the same manner as described previously for T2 and FLAIR; the resected region was traced using AFNI. The resulting lesion masks were mapped onto the surface at the gray-white junction using AFNI's 3dVol2Surf function.

MR Acquisition Protocol

All participants were scanned sagittally with a Philips Achieva 3T scanner in the NIH Clinical Center Radiology Department as follows: 1) 3D T1 weighted MPRAGE (T1): TR = 6.8–7.2, TE = 3.2 ms, TI = 900ms, flip angle = 90, voxel size = 0.75 x 0.75 x 0.8, acceleration factor 2 in slice direction, acquisition time = 7:02 min; 2) 3D T2 weighted FSE (T2): TR = 2500, TE = 225–245, voxel size = 1 x 1 x 2 or 1 x 1 x 1, acceleration factor 2 in slice and phase directions, acquisition time = 5:03 min; 3) 3D FLAIR: TR = 4800, TE = 271–415, TI = 1600, voxel size = 0.9 x 0.9 x 1, acceleration factor 2 in slice direction, 2.6 in phase direction, acquisition time = 6:10 min.

Image Preprocessing

For each individual subject, T2 and FLAIR images were co-registered to the T1wi with an affine transformation and a normalized mutual information cost function and sampled to the T1wi grid using Analysis of Functional NeuroImages software (AFNI\cite{Cox1996a}.  Registered images were visually assessed for alignment.  Cortical reconstruction was performed using T1w and T2w images as input to FreeSurfer's standard processing pipeline with  FreeSurfer v6.0.0  \cite{Dale1999,Fischl1999}.  Results were visually inspected and manually corrected if needed. Cortical surfaces were deformed to a standard mesh consisting of 36,002 vertices per hemisphere.  Intensity correction was implemented using an in-house procedure (see supplementary materials).