Prediction of seizure lateralization and postoperative outcome through the use of deep learning applied to multi-site MRI/DTI data: An ENIGMA-Epilepsy study

通过将深度学习应用于多部位 MRI/DTI 数据来预测癫痫偏侧化和术后结果：ENIGMA-癫痫研究

• 批准号: 9751025
• 负责人: Leonardo F Bonilha
• 金额: $ 44.43万
• 依托单位: UNIVERSITY OF CALIFORNIA, SAN DIEGO
• 依托单位国家: 美国
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-04-15 至 2022-03-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/9751025
• 关键词: Address; Affect; Age; Age of Onset; Algorithms; Antiepileptic Agents; Benchmarking; Capital; Characteristics; Classification; Clinical; Clinical Data; Complex; Country; Coupled; Data; Data Collection; Data Set; Databases; Development; Diagnostic; Diagnostic Procedure; Diffusion Magnetic Resonance Imaging; Disease; Electroencephalography; Epilepsy; Ethnic Origin; Evaluation; Geography; Gold; Grant; Image; Individual; Infrastructure; Institution; Lead; Left; Lesion; Lifting; Machine Learning; Magnetic Resonance Imaging; Methods; Modeling; Multimodal Imaging; National Institute of Neurological Disorders and Stroke; Network-based; Neurologic; Operative Surgical Procedures; Outcome; Partial Epilepsies; Patients; Pattern; Pharmacotherapy; Play; Population; Postoperative Period; Prediction of Response to Therapy; Probability; Reproducibility; Reproducibility of Results; Research; Resources; Role; Sample Size; Sampling; Scanning; Seizures; Site; Structure; Syndrome; Talents; Techniques; Temporal Lobe Epilepsy; Testing; Thinness; United States National Institutes of Health; Validation; base; biomarker identification; brain abnormalities; classification algorithm; cohort; computing resources; connectome; cost; cost effective; deep learning; design; gray matter; hands-on learning; imaging study; improved; innovation; interest; machine learning algorithm; nervous system disorder; neural network; neuroimaging; novel; novel strategies; personalized approach; sex; standard of care; surgery outcome; white matter

ABSTRACT Epilepsy is a devastating neurological illness that affects 65 million people worldwide. Approximately one-third of patients affected do not respond to antiepileptic drug therapy and require a thorough diagnostic work-up. Structural neuroimaging plays a pivotal role in the diagnostic evaluation of patients with focal epilepsy, identifying visible lesions in many patients that often coincide with the seizure focus. However, 20- 40% of patients have normal-appearing MRIs and this number appears to be growing. As a result, there is increased interest in identifying subtle gray and white matter network changes on non-invasive, quantitative MRI, including structural MRI (sMRI) and diffusion tensor imaging (DTI), that can help to delineate the epileptogenic network. Unfortunately, methods for selecting optimal features from sMRI/DTI data in patients with epilepsy that can address these clinical challenges have not been developed. There are at least two major barriers that have limited progress in this field. First, sample sizes have been insufficient to develop reliable classification algorithms in patients with focal epilepsy that lead to reproducible findings. The high cost of data collection - few studies scan more than 50-60 patients - has led to underpowered studies whose findings often fail to replicate and cannot adequately model confounds. Second, high computational demands have previously limited the feasibility of using sophisticated, feature-selection (i.e., Machine Learning; ML) algorithms in clinical settings. A new, large-scale data initiative (i.e., ENIGMA-epilepsy) acquired from 24 sites world-wide is now lifting these barriers and allowing for the development and validation of innovative data-driven approaches aimed at optimizing the use of MRI data in the evaluation of epilepsy. In this grant, we will leverage data collected through ENIGMA-Epilepsy—a new, cost-effective, innovative global approach that unblocks the power logjam by merging resources, data, capital infrastructure and talents of leading epilepsy centers from 14 countries across the world (2,149 patient and 1,727 healthy control MRI/DTI datasets). We will also leverage new developments in ML (i.e., deep learning) and network-based modeling (i.e., connectome- based approaches) and test whether these novel approaches improve upon classification accuracy relative to simpler, user-driven models. Our primary aim will be to test the ability of our deep learning approach (i.e., dense neural networks) to lateralize the seizure focus. In an exploratory aim, we will test the ability of our model to predict post-operative seizure outcomes. ENIGMA's harmonized approach will allow us to test our approach in over 24 datasets, diverse in age, ethnicity, age of onset, epilepsy duration, and surgical outcomes. This R-21 application addresses NIH's call for more reproducible studies by introducing a highly- powered design, and is directly aligned with NINDS's 2014 Epilepsy Benchmarks, which encourage the identification of biomarkers for assessing or predicting treatment response in patients with epilepsy.

摘要 癫痫是一种毁灭性的神经疾病，影响着全球6500万人。大致 三分之一的患者对抗癫痫药物治疗没有反应，需要进行彻底的诊断 体能训练。结构神经成像在局灶性脑脊髓炎的诊断评估中起着关键作用 癫痫，在许多患者中识别可见的病变，通常与癫痫灶一致。然而，20- 40%的患者核磁共振成像正常，而且这一数字似乎还在增长。因此，有 在非侵入性、量化的情况下识别细微的灰质和白质网络变化的兴趣增加 磁共振成像，包括结构磁共振成像(SMRI)和弥散张量成像(DTI)，可以帮助描绘 致痫网络。不幸的是，从患者的sMRI/DTI数据中选择最佳特征的方法 可以解决这些临床挑战的癫痫还没有开发出来。至少有两个 限制这一领域进展的主要障碍。首先，样本规模不足以进行开发 在局灶性癫痫患者中可靠的分类算法，导致可重复的发现。这个 数据收集的高成本-很少有研究扫描超过50-60名患者-导致研究动力不足 他们的发现往往无法复制，也不能充分模拟混乱。第二，计算量大 以前的需求限制了使用复杂的特征选择(即机器)的可行性 学习；ML)临床环境中的算法。 从全球24个地点获得的一项新的大规模数据倡议(即谜-癫痫)现在正在进行 消除这些障碍，并允许开发和验证创新的数据驱动方法 目的：优化MRI数据在癫痫评估中的应用。在这笔赠款中，我们将利用数据 通过谜-癫痫收集-一种新的、具有成本效益的创新全球方法，开启了 整合领先癫痫中心的资源、数据、资本基础设施和人才导致电力僵局 来自全球14个国家/地区(2,149名患者和1,727名健康对照的MRI/DTI数据集)。我们会 还可以利用ML(即深度学习)和基于网络的建模(即Connectome- 基于方法)，并测试这些新方法是否提高了相对于 更简单、用户驱动的模型。我们的主要目标是测试我们的深度学习方法的能力(即， 密集神经网络)，以确定癫痫发作的侧化焦点。在一个探索性的目标中，我们将测试我们的 用于预测术后癫痫结果的模型。Enigma的协调方法将允许我们测试我们的 方法在超过24个数据集，不同的年龄，种族，发病年龄，癫痫持续时间和手术结果。 这个R-21应用程序响应了美国国立卫生研究院对更多可重复性研究的呼吁，通过引入高度 动力设计，并直接与NINDS的2014年癫痫基准保持一致，这鼓励 用于评估或预测癫痫患者治疗反应的生物标志物的鉴定。