EEG Source Imaging of Epilepsy Sources by Means of Deep Neural Networks
Abstract number :
2.172
Submission category :
5. Neuro Imaging / 5B. Functional Imaging
Year :
2021
Submission ID :
1825558
Source :
www.aesnet.org
Presentation date :
12/1/2021 12:00:00 PM
Published date :
Nov 22, 2021, 06:43 AM
Authors :
Rui Sun, M.S. - Carnegie Mellon University; Abbas Sohrabpour, PhD – Carnegie Mellon University; Shuai Ye, MS – Carnegie Mellon University; Boney Josephs, M.B.B.S. – Mayo Clinic; Ben Brinkmann, PhDd – Mayo Clinic; Gregory Worrell, MD, PhD – Mayo Clinic; Bin He, PhD – Carnegie Mellon University
Rationale: Electroencephalography (EEG) is commonly used in the presurgical evaluation routine. Electrophysiological source imaging (ESI) is the process of estimating the underlying brain electrical activity from scalp E/MEG recordings and it has been widely used to help identify the epileptogenic zone by analyzing the interictal spikes or the ictal signals. Conventional ESI methods, however, require explicitly defining priors and regularization terms. This introduces complications such as tunning hyperparameters, making it challenging for the conventional ESI methods to provide the accuracy, efficiency, and objectivity needed for guiding treatment planning in epilepsy patients.
Methods: We have proposed a novel Source Imaging Framework using deep learning neural networks (SIFNet, https://www.biorxiv.org/content/10.1101/2020.05.11.089185v2), where synthetic data capable of realistically modeling brain activity and its corresponding EEG/MEG signals are employed to train a generalizable residual convolutional neural network to solve the ESI problem. The neural network contains five convolutional residual blocks followed by an average pooling layer and a fully connected layer. In the present study, we performed computer simulations to evaluate the performance of SIFNet, and evaluated it on the interictal spikes from 15 focal epilepsy patients who became seizure-free after surgery with at least 1 year follow-up. 21 ± 22 spikes were analyzed in each patient and compared to surgical resection outcome. Three conventional methods, sLORETA, LCMV, and dipole fitting were used as the benchmark to evaluate the SIFNET.
Results: SIFNet has a statistically significantly smaller localization error compared to sLORETA when evaluated on the simulation dataset (Fig 1A). We found that SIFNET’s performance was not affected by source depth and SNR contrary to all other. On the clinical dataset, SIFNet (3.68 ± 4.13 mm) shows superior performance for localizing interictal spike signals compared to the sLORETA, LCMV, and dipole fitting methods (6.13 ± 4.08, 9.22 ± 5.97, 12.99 ± 6.03 mm - Fig. 1B).
Conclusions: This novel data-driven source imaging algorithm, SIFNet, shows superior performance for localizing and imaging the epileptogenic tissue.
Funding: Please list any funding that was received in support of this abstract.: NIH R01NS096761; NIH R01EB021027.
Neuro Imaging