面向癫痫预测和灶点定位的多实例多标签深度学习模型研究

Automatic epileptogenic seizure prediction and focus localization based on computational techniques has great potential and is highly valuable to clinical applications. Seizure prediction is able to improve the treatment method for refractory epilepsy. Automatic focus localization provides an objective and precise focus result, which reduces the work of clinical doctors. By deeply considering the property of the two problems, this project defines the focus localization as a multi-label multi-class classification problem and defines the seizure prediction as a multi-instance learning problem. This is a new viewpoint and enables the supervised learning for the focus localization problem. Based on these abstractions, we are going to use deep learning as a basic framework to delve the intrinsic patterns representing foci and seizures. The project uses scalp Electroencephalograph (sEEG) and intracranial EEG (iEEG) as input and studies the following problems. First, a mathematical model for multi-label multi-class classification will be built for focus localization. The model will be combined with deep neural network for practical use. Then, for seizure prediction, we are going to find a multi-instance learning formulation, which can be integrated with recurrent neural network. After performing the two models, we want to study the relationship between the features of seizure prediction and focus localization. The underlying idea is that both features represent the abnormal pattern of intracranial EEG, although they can be different in the magnitude. So it is highly possible they can help each other. We mainly research the mathematical model to connect the two objectives and how to develop a deep algorithm to achieve joint learning and optimization, in order to extract the joint feature. In addition, we will develop an automatic system for the seizure prediction and focus localization using the techniques developed above.

基于信息计算技术的癫痫预测和灶点自动定位有望改善癫痫诊疗手段、为临床提供一个客观参考、提高诊断准确率，从而具有重要的研究和临床应用价值。本项目紧扣癫痫预测和灶点定位问题的特点，从全新角度将灶点定位抽象为多标签多类分类问题，同时将癫痫预测抽象为多实例学习问题。基于这两种问题抽象，本项目以深度学习为基本框架，试图挖掘真正表征灶点和癫痫发作的内在模式特征。项目以脑电(sEEG和iEEG)作为输入，研究面向癫痫定灶的多标签多类分类数学模型和对应的深度神经网络；研究面向癫痫预测的多实例学习模型及相关深度神经网络。在此基础上，进一步研究基于iEEG的癫痫预测和灶点定位在脑电特征模态上的关联，潜在的想法是预测或灶点定位均基于表示癫痫发作的异常脑电模式，这两种特征应该是互助的。重点考虑勾联两个目标的数学模型及计算模型、及如何进行联合优化，以实现预测和定位联合特征提取。最后拟开发一个预测与灶点定位系统。

本项目的研究目标为建立基于多标签多类分类的灶点定位方法和基于多实例学习的癫痫预测方法，解决基于脑电的癫痫灶点定位和基于颅内脑电的癫痫预测。围绕上述研究目标，本项目重点研究了计算机辅助的自动化癫痫预测和癫痫灶点定位方法，主要研究成果包括：1）癫痫数据库方面，建立了浙大癫痫数据库，共采集癫痫病例100人、总时长超过2.6万小时的大规模癫痫数据库，完成部分标注并公开，为相关研究提供了良好的数据支撑；2）癫痫预测和癫痫脑信号特征提取方面，提出了多实例多尺度卷积神经网络、图卷积神经网络等多种癫痫特征提取方法，有效解决癫痫发作前期标签不准确、脑信号特征微弱条件下有效特征提取问题，提升了癫痫预测效果；3）自动化癫痫灶点标定方面，建立了基于循环神经网络-格兰杰因果检验的脑连通性分析方法、基于黎曼流形的癫痫状态自动划分和灶点自动定位方法等一系列方法，实现了计算机辅助全自动癫痫灶点定位，并已形成系统与医院合作进行临床测试。总体而言，项目预期核心指标均符合预期或超出预期完成，项目按照研究计划执行顺利。