Towards personalized multimodal lesion mapping of temporal lobe epilepsy: a multisite validation
Poster No:
1474
Submission Type:
Abstract Submission
Authors:
Judy Chen1, Jessica Royer1, Oualid Benkarim2, Alexander Ngo2, Ella Sahlas3, Sara Larivière4, Raúl Rodriguez-Cruces2, Ke Xie3, Yifei Weng5, Jordan DeKraker6, Andrea Bernasconi1, Neda Bernasconi1, Dewi Schrader7, Zhiqiang Zhang8, Luis Concha9, Boris Bernhardt2
Institutions:
1Montreal Neurological Institute and Hospital, Montreal, QC, 2Montreal Neurological Institute and Hospital, Montreal, Quebec, 3McGill University, Montreal, Quebec, 4Brigham and Women’s Hospital, Portsmouth, NH, 5Nanjing University, Nanjing, Nanjing, 6McGill University, Montreal, Canada, 7University of British Columbia, Vancouver, BC, 8Jinling Hospital, the First School of Clinical Medicine, Southern Medical University, Nanjing, Jiangsu, 9Universidad Nacional Autonoma de Mexico, Queretaro, Please select an option below
First Author:
Co-Author(s):
Zhiqiang Zhang
Jinling Hospital, the First School of Clinical Medicine, Southern Medical University
Nanjing, Jiangsu
Jinling Hospital, the First School of Clinical Medicine, Southern Medical University
Nanjing, Jiangsu
Introduction:
Drug-resistant epilepsy affects over 20 million people worldwide, and a majority of these patients are diagnosed with temporal lobe epilepsy (TLE)1. Diagnostic methods to lateralize the lesion for surgical resection remains a challenge due to the heterogeneity in clinical phenotypes. Past work leveraging MRI to individually characterize patients and lateralize their pathology has been promising but remains limited to a single modality and/or single site2. Here, we leverage our open-source image processing and analysis tools to: 1) map the distribution of patient-specific alterations in neocortical and hippocampal subregions from structural (T1w) and diffusion (DTI) scans across three sites; 2) evaluate the efficacy of these features to lateralize patients in a machine learning framework.
Methods:
Participants. We studied 388 individuals aggregated from 3 independent sites. Our cohort comprised 123 TLE patients (61 females; mean±SD: 32.4±11.1 years; 57 left-sided focus) and 265 healthy individuals (94 female; mean±SD age: 29.0±9.6 years). All participants underwent 3T T1-weighted (T1w) and diffusion MRI scans. The Nanjing cohort consists of 122 healthy controls (HCs) and 70 TLE patients; EpiC cohort consists of 32 HCs and 18 TLE patients; and MNI cohort consists of 111 HCs and 35 TLE patients.
Image processing and analysis. Multimodal imaging data were processed using micapipe_v0.2.03 and hippunfold_v1.3.04. Post-processing using z-brains (https://github.com/MICA-MNI/z-brains) involved z-scoring surface based features in the neocortex and hippocampus of each patient against their site-matched healthy controls while controlling for age and sex in fsLR 64k vertex space and 5k hippocampal space. Right-TLE patients were flipped to reflect ipsilateral and contralateral regions with respect to seizure focus. Significantly altered vertices were defined at |z| 1.96 and the probability of each vertex being altered across the patient cohort was mapped across three features: cortical thickness (CT), fractional anisotropy (FA), and apparent diffusion coefficient (ADC). Logistic regression (LR) and support vector machine (SVM) algorithms were trained with 10-fold cross validation on 5k fsLR vertex-wise neocortical and hippocampal data to lateralize the seizure focus in TLE.
Image processing and analysis. Multimodal imaging data were processed using micapipe_v0.2.03 and hippunfold_v1.3.04. Post-processing using z-brains (https://github.com/MICA-MNI/z-brains) involved z-scoring surface based features in the neocortex and hippocampus of each patient against their site-matched healthy controls while controlling for age and sex in fsLR 64k vertex space and 5k hippocampal space. Right-TLE patients were flipped to reflect ipsilateral and contralateral regions with respect to seizure focus. Significantly altered vertices were defined at |z| 1.96 and the probability of each vertex being altered across the patient cohort was mapped across three features: cortical thickness (CT), fractional anisotropy (FA), and apparent diffusion coefficient (ADC). Logistic regression (LR) and support vector machine (SVM) algorithms were trained with 10-fold cross validation on 5k fsLR vertex-wise neocortical and hippocampal data to lateralize the seizure focus in TLE.
Results:
Proportion maps. The temporal lobe demonstrated the highest proportion of neocortical alterations across all features (thickness, FA, and ADC), and were most pronounced with ADC, reaching over 40% (F1A). However, mean proportion of each region parsed through the Desikan-killiany atlas revealed high neocortical variability remains even in the best performing feature (ADC) as no specific region was altered in more than 50% of patients. Hippocampal maps showed moderate atrophy, subtle FA decreases, and highly pronounced ADC signal (F1A). An anterior-posterior gradient was most prominent across the ADC feature and somewhat noticeable in the hippocampal atrophy patterns. Site-specific alteration proportion maps for both neocortex and hippocampus demonstrated preservation of the overall neocortical alteration patterns across sites (F2B).
Lateralization performance. Classifiers trained on hippocampal data vastly outperformed those on neocortical data, achieving the highest accuracy score of 85.4% (logistic regression on ADC) as compared to the maximum accuracy score of 70.8% (logistic regression on ADC & FA). Overall highest accuracy score of 87.1% was SVM trained on ADC&CT neocortical and hippocampal data.
Lateralization performance. Classifiers trained on hippocampal data vastly outperformed those on neocortical data, achieving the highest accuracy score of 85.4% (logistic regression on ADC) as compared to the maximum accuracy score of 70.8% (logistic regression on ADC & FA). Overall highest accuracy score of 87.1% was SVM trained on ADC&CT neocortical and hippocampal data.
·Figure 1. Mapping neocortical alteration probability maps. A | Hippocampal alteration maps show moderately distributed hippocampal atrophy, subtle FA decreases, and highly pronounced ADC signal. Tempo
·Figure 1. Lateralization accuracy scores of machine learning algorithms trained on neocortical and hippocampal data. Logistic regression (LR) and support vector machine (SVM) algorithms (10-fold cross
Conclusions:
MRI phenotypes among patients are too variable to rely on single modality for accurate lesion lateralization. It is possible to accurately lateralize TLE patients using personalized lesion mapping techniques with machine learning classification models. We present here a multisite validation of this technique, achieving a maximum accuracy score of 87.1%.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Diffusion MRI Modeling and Analysis
Multivariate Approaches
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping
Keywords:
ADULTS
Epilepsy
Machine Learning
MRI
Open-Source Software
STRUCTURAL MRI
White Matter
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
2. Sone et al., 2021. Front Neurosci, 10.3389/fnins.2021.684825.
3. Cruces, R et al., 2022 NeuroImage, 10.1016/j.neuroimage.2022.119612.
4. DeKraker, J et al., 2022 eLife, 10.7554/eLife.77945.