【作者】 李鑫；

【导师】 Jean Louis COATRIEUX； 杨淳沨； 刘庭华；

【作者基本信息】 东南大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 作为一种临床上常见的神经性疾病,癫痫是由于大脑神经元异常放电引起的中枢神经系统功能失调。虽然大部分癫痫患者可以通过药物进行治疗,但是仍然约有三分之一的癫痫患者无法通过药物抑制发作。针对该类患者,医生可以采用外科手术切除癫痫病灶区域来进行治疗。手术成功的关键在于术前能否准确定位出癫痫病灶区域。但由于癫痫发作时,异常信号的快速传播导致很难准确定位到癫痫病灶区域。因此,本文致力于基于脑电信号的癫痫病灶区域定位算法研究,该研究可以为术前癫痫病灶切除手术时提供有力诊断依据。图论分析可以对癫痫发作时不同区域之间连接构建的脑网络的病理性特征评估来定位癫痫病灶区域。先前的研究使用基于卡尔曼滤波器的自适应定向传递函数计算时变连通性来构建脑网络,但是卡尔曼滤波器的性能依赖于自适应常数,如果自适应常数设置错误,可能导致无法准确构建脑网络。为了克服这一缺点,本文提出自校正定向传递函数结合图论分析的新算法。该算法基于自校正优化卡尔曼滤波器算法估计时变自回归系数,根据估计的系数来计算连通性矩阵,使用图论分析方法衡量每个节点在网络中的图论特征来定位癫痫病灶区域。该算法的主要优势是可以更加准确估计大脑不同区域之间时变连接。实验结果表明,在发作早期阶段该算法预测的癫痫病灶区域和实际手术切除区域较为重合,能够为癫痫术前评估提供有效手段。与图论分析不同,网络动力学模型是通过对癫痫发作等病理状态转变的评估来定位癫痫病灶区域。目前很多研究使用功能连通性来构建网络动力学模型开发癫痫病灶定位算法,但是功能连通性无法体现癫痫发作时大脑不同区域之间连接的方向性,不完整的方向性信息可能会导致定位癫痫病灶区域产生偏差。为了克服这一缺点,本文提出基于效应连通性的网络动力学模型新算法,该算法基于格兰杰因果构建网络动力学模型,衡量每个节点对网络产生癫痫发作的重要程度来定位癫痫病灶区域。该算法的主要优势在于它融合了大脑不同区域之间定向连接,可以更准确地描述神经网络的动态特征和传播机制。实验结果表明,本文提出的算法比现有方法能更准确地定位癫痫病灶区域,该算法可以为临床医生精确定位癫痫病灶区域提供客观依据。定位癫痫病灶区域是治疗耐药性癫痫患者的关键。本文分别从图论分析和网络动力学模型方向上提出了两种算法来定位癫痫病灶区域,在耐药性癫痫患者的数据上验证了两种算法的有效性,可以为癫痫术前评估提供有效手段。更多还原

【Abstract】 As a common clinical neurological disorder,epilepsy is a dysfunction of the central nervous system caused by abnormal neuronal discharges in the brain.Although most patients with epilepsy can be treated with medications,there are still about one-third of patients with epilepsy whose seizures cannot be suppressed with medications.For this group of patients,doctors can treat them by surgically removing the area of the epileptic lesion.The key to successful surgery is the ability to accurately locate the epileptic foci before surgery.However,the rapid propagation of abnormal signals during seizures makes it difficult to accurately locate the epileptic focal area.Therefore,this thesis is dedicated to investigating the localization algorithm of the epileptogenic zone based on EEG signals,which can provide a strong diagnostic basis for preoperative epileptic lesion resection surgery.Graph theoretic analysis allows the evaluation of pathological features of brain networks constructed by connecting different regions during seizures to localize epileptic focal regions.Previous studies have used Adaptive Directed Transfer Function based on the Kalman filter to calculate time-varying connectivity to construct brain networks,but the performance of the Kalman filter depends on the adaptive constant,and if the adaptive constant is set incorrectly,it may lead to inaccurate construction of brain networks.To overcome this drawback,this thesis proposes a new algorithm of Self-Tuning Directed Transfer Function combined with graph theory analysis.The algorithm is based on the Self-Tuning Optimized Kalman filter algorithm to estimate the time-varying autoregressive coefficients,calculate the connectivity matrix based on the estimated coefficients,and use graph theoretic analysis to measure the graph theoretic features of each node in the network to locate the epileptogenic zone.The main advantage of this algorithm is that it allows more accurate estimation of time-varying connectivity between different regions of the brain.Experimental results show that the predicted epileptogenic zone by the algorithm and the actual surgical resection areas overlap more in the early stages of seizures,which can provide an effective means for preoperative assessment of epilepsy.Unlike the graph theoretic analysis,dynamical network models are used to localize epileptogenic zone through the assessment of pathological state transitions such as seizures.Many studies have used functional connectivity to construct dynamical network models to develop epileptogenic zone localization algorithms,but functional connectivity cannot reflect the directionality of connections between different regions of the brain during seizures,and incomplete directional information may lead to bias in localizing epileptogenic zone.To overcome this drawback,this thesis proposes a new algorithm for dynamical network model based on effective connectivity,which constructs a dynamical network model based on Granger causality and measures the importance of each node to the network in generating seizures to localize the epileptogenic zone.The main advantage of this algorithm is that it incorporates directed connections between different regions of the brain,which can more accurately characterize the dynamics and propagation mechanisms of neural networks.Experimental results show that the algorithm proposed in this thesis can localize the epileptogenic zone more accurately than existing methods,and the algorithm can provide an objective basis for clinicians to precisely locate the epileptogenic zone.Locating the epileptogenic zone is the key to treating patients with drug-resistant epilepsy.In this thesis,we propose two algorithms to localize epileptogenic zone in the direction of graph theory analysis and network dynamics model,respectively,and validate the effectiveness of both algorithms on data from patients with drug-resistant epilepsy,which can provide an effective means for preoperative evaluation of epilepsy.更多还原