Exploring Morphological Characterization of MRI-negative Child Epilepsy Based on Machine Learning

355 

Abstract Submission 

Kai Zhang¹, Jinping Xu², Fan Xinxin³, Zhewei Zhang⁴, Xiaodong Zhang⁵

¹Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, GuangDong, ²Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangzhou, ³Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, ⁴Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, ⁵Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences, Shenzhen, Guangdong

Kai Zhang  
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Shenzhen, GuangDong

Jinping Xu  
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Shenzhen, Guangzhou

Fan Xinxin  
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Shenzhen, Guangdong

Zhewei Zhang  
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences
Shenzhen, Guangdong

Xiaodong Zhang  
Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences
Shenzhen, Guangdong

FCD is a disease caused by abnormal proliferation and differentiation of focal cortical neurons, cortical architecture and migration, which is the most common cause of refractory secondary epilepsy in children.In epilepsy surgery, FCD accounts for about 40% to 50% of pediatric epilepsy surgery patients.
Compared with adults, children's brains are still in the developmental stage, and the morphological characteristics of lesions are more subtle and difficult to distinguish. At the same time, there are few sample data for children with epilepsy, and traditional research methods based on regions or individuals are used. Therefore, for these patients, we propose a machine learning method based on brain surface vertices, which deeply utilizes the morphological characteristics of MRI-negative children with epilepsy and improves the Predictive accuracy in MRI-negative children with epilepsy and provides a theoretical basis. Provide assistance to doctors in clinical diagnosis and treatment.

A MRI-positive cohort of 72 patients from publicly available data(https://www.nature.com/articles/s41597-023-02386-7). The other one MRI-positive cohort of 11 child patients and MRI-negative cohort of 8 child patients from Shenzhen Children's Hospital following permission by the hospital ethical review board. Patients younger than 3 years of age or Prognosis is not standard(<Engel class II) were excluded.A control group of 32 participants with no history of any neurological diagnosis from publicly available data ABIDE.
All patients and controls were scanned on a 3-T MRI system.3D structural T1w images using the following protocols
[TR] = 2300 milliseconds,[TE] = 2.74 milliseconds,[FOV]= 256 × 256 mm,[FA] = 8°,voxel size=1×1×1 mm3.Cortical reconstructions were generated using FreeSurfer version 7.3(https://surfer.nmr.mgh.harvard.edu/) for all participants.Lesion masks were manually delineated for the all patients by an experienced pediatric neuroradiologist. The lesion masks were first mapped onto the individual surface reconstructions and then onto the bilaterally symmetric template(fsaverage_sym).Measures of morphological/intensity features. The following measures were calculated per vertex across the cortical surface in all participants: cortical thickness,intensity at the gray-white matter contrast,curvature,sulcal depth,intrinsic curvature and boundary sharpness coefficient(BSC).We evaluated the discrimination of BSC.All features were smoothed with a 10mm Gaussian kernel and underwent two normalization procedures:within-subject z scoring and a between-subject z scoring.And then,all features were registered to fsaverage_sym : a bilaterally symmetrical template space. The Scikit-learn toolbox was used to create a linear classifier that classifies each vertex as diseased or non-diseased.The classifier is trained using data from MRI-positive cohort.And then input features from MRI-negative cohort per-vertex and output predictions. grouped into neighbor-connected clusters of vertices.If the predicted vertex cluster hits the manually outlined ROI, it is considered a successful detection.

The network was trained using a leave-one-out crossvalidation approach to assess the accuracy of the classifier on the MRI-positive cohort . And then trained on all 83 MRI-positive and tested on MRI-negative cohort.
Of the 83 patients with visible FCD on MRI, the classifier was able to detect the lesion in 67 (sensitivity = 81%). In 5 of the 8 patients (sensitivity = 62%) with negative MRI, the predicted vertex cluster overlapped with the hand-delineated lesion mask.No clusters were detected In healthy controls(specificity = 100%).

We propose an automatic method for detecting FCD and test the feasibility of BSC in predicting FCD lesions. After adding BSC feature, the sensitivity of vertex classification is increased by 4%. Ultimately, there was 81% sensitivity in the positive MRI cohort and 62% sensitivity in the negative MRI cohort, with a specificity of 100%.

Non-invasive Magnetic/TMS

Neurodevelopmental/ Early Life (eg. ADHD, autism) ¹

Classification and Predictive Modeling ²

Epilepsy

Machine Learning

Pediatric Disorders

STRUCTURAL MRI