Combining Traditional MRI and a Novel Outlier Detection Method to Increase Diagnostic Yield for Surgical Epilepsy. (16th November 2020)

• Record Type: Journal Article
• Authors: Talati, Pratik; Spader, Heather; O'Muircheartaigh, Jonathan; Stone, Scellig S.D; Bolton, Jeffrey; Prabhu, Sanjay
• Is Part Of: Neurosurgery. Volume 67(2010)Supplement 1
• Publication Date: 2020-11-16
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyaa447_662
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Children with epilepsy often have subtle focal anomalies on brain MRI that can be challenging to identify based on their varied location. Image based outlier detection can highlight these differences without strong underlying hypotheses. METHODS: We collected clinical T1-weighted and T2-weighted FLAIR images acquired on a 3T scanner from 35 pediatric patients with focal epilepsy at Boston Children's Hospital after IRB approval. Nonlinear registration was used to register each patient's structural and functional image to a pediatric brain template (Avants et al, 2014). Image intensity was normalized per volume to a zero mean and unit standard deviation. Each patient's outlier "zeta" score was calculated at a voxelwise level based on the voxel intensity and its neighbors (Mah et al, 2014). The ILAE seizure outcome classification was used with scores above two considered good. If the MRI predicted the resection area and the ILAE score was good, seizure outcome was correctly predicted. If the seizure outcome was bad and not predicted by MRI, seizure outcome was also correctly predicted. Discordance was considered a fail. RESULTS: The 35 patients (22 male) had an average age of 42 months (range: 6 to 144 months); 27 patients (77%) had a good outcome. The brain lobe where the epilepsy resection occurred was localized 63% of the time with the outlier detection method; MRI alone (from the original report) localized 60% of the lesions, and MRI combined with outlier detection localized 80% of the lesions. The seizure outcome was correctly predicted in 54% of cases using outlier detection, 43% with MRI alone, and 71% with traditional MRI and outlier detection combined. CONCLUSION: Using MRI plus outlier detection, we were able to localize the area of seizure resection approximately 80% of the time and predict seizure outcome 71% of the time. Outlier detection may act as a simple adjunct to help localize seizure foci in large heterogeneous populations as seen in childhood epilepsy case series.