Imaging Epilepsy Sources with Biophysically Constrained Deep Neural Networks

项目摘要 本项目的目标是开发和验证一种新的电生理源成像（ESI）方法 基于生物病理学约束的深度神经网络（BioDNN），显著改善手术规划， 耐药局灶性癫痫患者。癫痫影响着全世界约7000万人。约33% 在340万美国癫痫患者中，癫痫发作并不是单靠药物就能控制的。癫痫手术是 治疗耐药性局灶性癫痫的最可行的选择，只有当癫痫发作源可以准确定位 并安全转移临床上需要创新技术工具，以更好地制定病灶切除术的手术计划。 癫痫我们在这个项目中提出了一种新的ESI技术，该技术基于生物制药学约束的深度神经网络， 网络（BioDNN）提供准确，可靠和客观的时空估计的基础 致痫区（EZ）。创新之处在于，经过训练的神经网络，能够对大脑来源进行成像 而不需要由操作员针对每个新的数据实例来调整模型的超参数，因此 使得该技术客观且易于在临床环境中使用。我们的具体目标是：目标1。建立 验证BioDNN用于从EEG发作间期癫痫样放电（IED）成像癫痫组织 局灶性癫痫患者我们将建立，优化和验证拟议的BioDNN成像EZ， 200例局灶性耐药癫痫（DRE）患者EEG中的IED，与临床“地面实况”进行比较。目标2. 脑磁图发作间期癫痫样发作致痫组织成像BioDNN模型的建立与验证 局灶性癫痫患者的出院。我们将开发和优化BioDNN模型，用于从 MEG-BioDNN模型在80例局灶性DRE中的应用 患者与临床“地面实况”的比较。目标3。开发和验证BioDNN成像模型 局灶性癫痫患者发作期脑电图中的致痫组织我们将开发BioDNN， SOZ从头皮发作脑电图，并验证它从高密度发作脑电图记录在120局灶性DRE患者， 与临床“基本事实”进行比较。成功完成拟议的研究将建立一个新的 利用机器学习技术对发作间期的潜在致痫组织进行非侵入性定位和成像 和发作的电生理学生物标志物。这种新技术的建立有望大大提高 提高颅内EEG电极植入的精度，并辅助手术计划， 手术结果的改善，并使许多耐药癫痫患者受益。 1

Project Summary The goal of this project is to develop and validate a novel electrophysiological source imaging (ESI) approach based on biophysically constrained deep neural networks (BioDNN), to significantly improve surgical planning in drug resistant focal epilepsy patients. Epilepsy affects about 70 million people worldwide. For approximately 33% of the 3.4 million Americans with epilepsy, seizures are not controlled by medications alone. Epilepsy surgery is the most viable option for curing drug resistant focal epilepsy, only if seizure sources can be accurately localized and safely removed. There is a clinical need to innovate technological tools for better surgical planning of focal epilepsy. We propose in this project a novel ESI technology based on biophysically constrained deep neural network (BioDNN) to provide accurate, robust, and objective spatio-temporal estimates of the underlying epileptogenic zone (EZ). Of innovation is that the trained neural network, is capable of imaging brain sources without the need to tune the model’s hyper-parameters by an operator for every new instance of data, thus making the technique objective and easy-to-use in clinical settings. Our specific aims are: Aim 1. Establishing and Validating the BioDNN for Imaging Epileptogenic Tissue from EEG Inter-ictal Epileptiform Discharges (IEDs) of Focal Epilepsy Patients. We will establish, optimize and validate the proposed BioDNN for imaging EZ from IEDs in EEG in 200 focal drug resistant epilepsy (DRE) patients, in comparison to clinical “ground truth". Aim 2. Developing and Validating the BioDNN Model for Imaging Epileptogenic Tissue from MEG Inter-ictal Epileptiform Discharges of Focal Epilepsy Patients. We will develop and optimize the BioDNN model for imaging EZ from MEG IEDs and validate the MEG-BioDNN model and compare with the EEG-BioDNN model in 80 focal DRE patients in comparison to clinical “ground truth. Aim 3. Developing and Validating the BioDNN Model for Imaging Epileptogenic Tissue from Ictal EEG of Focal Epilepsy Patients. We will develop the BioDNN for imaging the SOZ from scalp ictal EEG and validate it from high density ictal EEG recordings in 120 focal DRE patients, in comparison to clinical “ground truth”. The successful completion of the proposed research will establish a novel machine learning technology to non-invasively localize and image underlying epileptogenic tissue from interictal and ictal electrophysiological biomarkers. The establishment of such a novel technology promises to significantly improve the precision of intracranial EEG electrodes implantation and aid surgical planning, leading to significant improvement in surgical outcomes, and benefiting numerous drug resistant epilepsy patients. 1