Automated Identification of Focal Cortical Dysplasia on Structural MRI Using a Novel Machine Learning Approach
Abstract number :
1.253
Submission category :
5. Neuro Imaging / 5A. Structural Imaging
Year :
2019
Submission ID :
2421248
Source :
www.aesnet.org
Presentation date :
12/7/2019 6:00:00 PM
Published date :
Nov 25, 2019, 12:14 PM
Authors :
Kathryn Snyder, NIH; Souheil J. Inati; Emily P. Whitehead, Self; William H. Theodore, NIH; Kareem A. Zaghloul, NIH; Sara K. Inati, NIH
Rationale: Identification of epileptogenic lesions plays an important role in presurgical evaluation of patients with epilepsy. Identification of focal cortical dysplasias is a particular challenge, as subtle lesions can be overlooked. In these cases, quantitative analysis can aid in lesion identification, but it often requires significant manual intervention or large training sets. Here, we implement a lesion detection algorithm using a cosine similarity score to assess lesion likelihood for each patient and to improve identification of FCDs. Methods: MRIs were obtained from 30 healthy volunteers (12 females, ages 8-63) and 15 patients with focal cortical dysplasia (11 females, ages 15-53). Study data were acquired on NIH Clinical Center 3T Philips Achieva MRI Scanners. Imaging included 3D volumetric images (MPRAGE, T2, and FLAIR). Images were coregistered to the T1, and a subject specific intensity normalization procedure was carried out for each contrast through local normalization by the signal mean. Surface reconstruction and cortical surface metrics were calculated using FreeSurfer. A set of 39 multiscale 3D local image statistics were computed for three MRI contrasts and ascribed to each node on the gray-white surface. Each feature was standardized across the whole cortex within each subject. Principal component analysis fit to all cortical vertices for all controls was applied to the resulting feature set, and 14 components were kept (90% explained variance). A non-linear transformation followed by PCA was applied to the resulting components iteratively so that the feature set followed a multivariate normal distribution. Lesion masks were manually defined and projected onto the GWJ. The average feature vector of all FCD nodes within MRI positive patients was computed and unit normalized. A weighted cosine similarity measurement was calculated for each subject by computing the dot product of the unit normalized averaged FCD feature vector with the feature vector at each cortical node. A local normalization procedure was carried out to account for normal variability in the resulting cosine map, where the mean and standard deviation at each node was calculated as the smoothed (FWHM 20) average of the cosine map in controls. The resulting map was then thresholded (Z > 3) and clustered with a minimum cluster size of 30 nodes. Results: Fifteen patients with FCDs were studied. High similarity scores were observed in 13/15 patients. Additionally, FCDs were detected in 12/15 patients (10/10 MRI positive patients; 2/5 MRI negative patients), where detection was defined as the colocalization of a cluster with the manual lesion label. Of the 30 controls studied, no clusters were detected (AUC = 1). Prior to local normalization, increased similarity scores were observed in some normal cortical regions primarily consisting of the insula and the temporal pole. Conclusions: The results of this study indicate that postprocessing using local image statistics can be used to evaluate lesion likelihood for individual patients, and approaches such as this one that require only small training sets using publicly available tools may facilitate more widespread adoption of these techniques into clinical practice. Additionally, automated detection of focal cortical dysplasia can be used to guide surgical resection or implantation strategies in patients with drug resistant focal epilepsy. Funding: This work was supported by the NIH Intramural Research Program.
Neuro Imaging