【作者】 陈文静；

【导师】 李真林；

【作者基本信息】 四川大学 ， 医学技术（影像）， 2022， 硕士

【摘要】 背景:癫痫(epilepsy)是一种临床常见的中枢神经系统疾病。其中30%的癫痫患者抗癫痫药物治疗无效,最终发展成难治性癫痫,通常这部分患者最好进行手术切除或者离断致痫脑区。因此术前精准定位致痫灶是手术成功的关键。在术前评估中,颅内脑电(包括皮层脑电(electrocorticogram,ECoG)及立体脑电(stereoelectroencephalogram,SEEG))技术扮演着至关重要的角色。在ECoG方面,过去二十年已有大量研究提出脑电信号中的高频振荡信号(highfrequency oscillations,HFOs)是一种新的生物标志物,可用于定位癫痫发作的起始区域(seizure onset zone,SOZ),但目前人工分析方法被认作是HFOs的金标准,存在严重主观性,容易将高频干扰噪声和多种癫痫样尖峰波形等两类信号误检为HFOs信号,此外还极度耗时。在SEEG方面,作为重要检测手段的立体电刺激(electric stimulation,ES)在国内外各个医疗中心之间至今未实现技术统一和标准化,尤其是目前的一些研究结果显示ES的多种结果与癫痫手术预后相关性的研究结论不一致。目的:在ECoG方面,提出一套基于机器学习方法的HFOs自动检测算法及其用于致痫灶定位的方法;在SEEG方面,采用统计学分析围绕多因素与癫痫手术预后的相关性展开,重点研究通过ES的多种结果定位致痫灶与良好预后的相关性。初步探究ECoG的HFOs自动检测技术和SEEG的ES结果与预后相关性统计学研究在癫痫致痫灶术前定位中的临床应用价值,以期更好地实现精准定位癫痫致痫灶的目标。方法:回顾性纳入我院神经外科2015年9月至2021年8月的所有颅内电极植入患者40例,包含皮层电极植入术的ECoG患者10例和立体定向电极植入术的SEEG患者30例。全部患者都是抗癫痫药物治疗无效并且严重影响患者生活质量的难治性癫痫,术前经过完整的无创评估方法(包括视频脑电图VEEG、MRI、CT、PET-CT及神经心理学检查等)仍不能进行致痫区定侧或定位,术后接受短期、中期及长期随访,依据随访情况按Engel分级进行Ⅰ-Ⅳ级分类。收集所有患者的详细临床资料和数据,包含癫痫发作病史、发作症状、发作频率、发作时长、视频脑电图、神经心理学检查及相关影像学检查等结果,构建一套单中心、小样本的癫痫患者颅内皮层脑电ECoG数据库和一套颅内立体脑电SEEG数据库。在ECoG方面,首先将检测到的ECoG使用小波包滤波预处理,再使用误差反向传播神经网络(Back propagation,BP)进行特征提取和自动分类检测出HFOs,然后根据检测出的HFOs高发区确定相应的癫痫灶位置;最后,对HFOs在各个脑电信号通道的出现率进行了统计,获得的HFOs在脑电导联的所在区域,对比医师通过解剖-电-临床及随访预后情况综合分析确定的致痫灶,评估了该方法HFOs的检测性能及致痫灶定位的准确度。在SEEG方面,首先基于影像建模的引导下精确ES,其次利用统计学方法将相关因素与预后进行多变量分析,分析的变量包括PET、头皮脑电、定侧、起源区定位、植入部位、是否手术切除、病理结果、发病年龄、电极根数、随访时间及高频、低频电刺激结果以及患者年龄、性别;最后将统计结果进一步进行两两比较,采用Bonferroni法调整α水平。结果:在ECoG方面,本研究纳入3例ECoG脑电监测患者(男性2例,女性1例,年龄分布在19-26岁,平均年龄21.67±3.09岁),由脑电图和神经外科医师人工标记出的HFOs共计2790个,本文采用的机器学习方法能准确地检出其中的2511个HFOs,误检共计208个,结果灵敏度是90%,误检率是7.46%。其中,2例局灶性癫痫患者HFOs高发的脑电通道区与临床医师综合评估勾画出的癫痫灶区域有较大的重叠或非常靠近临床诊断的癫痫灶;另外1例多灶性癫痫患者HFOs高发的通道相对比较分散,其预估的病灶区且与医生所勾画的癫痫灶区域重合区域相对比较小。在SEEG方面,本研究纳入25例SEEG脑电监测患者(男性15例,女性10例,年龄分布在3-37岁,平均年龄22.56±8.33岁),本研究将电刺激结果分为先兆、惯常发作、其他三组。患者的预后以发作与否为标准,发作与否中对应的“是”指无癫痫复发,代表预后良好(指近远期随访为Engel分级I);发作与否中对应的“否”指有癫痫复发,代表预后不佳(指近远期随访为Engel分级Ⅱ-Ⅳ)。结果发现预后良好的19例患者中LFS结果包括14例先兆患者,4例惯常发作患者,1例其他患者(未刺激出先兆及惯常发作)。多变量分析统计结果显示除低频电刺激结果外,其他变量各组间与是否发作均没有统计学差异(P>0.05)。低频电刺激结果显示,先兆组患者中没有出现癫痫发作,惯常发作组中有1位患者发生癫痫发作(20.0%),而其他组中有5位患者发生癫痫发作(83.3%)。Fisher精确检验结果显示,三组差异具有统计学意义(χ~2=14.273,P<0.001)。进一步进行两两比较,采用Bonferroni法调整α水平的结果显示:低频电刺激结果中惯常发作组和其他组、先兆组和惯常发作组与随访患者癫痫是否发作无统计学差异(Bonferroni校正,P>0.0167);而先兆组和其他组与随访患者癫痫是否发作有统计学差异(Bonferroni校正,P<0.0167)。结论:在ECoG方面,本文初步研究结果显示采用机器学习方法可有效检测出皮层脑电ECoG中的HFOs并可根据检测出的HFOs高发区域自动定位相应的致痫灶区域,该区域与临床医师诊断的致痫灶区域(尤其是局灶性癫痫)有很高的相关性;同时该方法仅需要患者15分钟发作间期的脑电监测数据,在很大程度上将大大减轻患者长时间监测的痛苦和临床脑电医师人工读图的工作量。在SEEG方面,本文基于ES诱发脑电,围绕患者的多因素与预后相关性研究结果显示,除低频电刺激结果外,其他变量各组间与是否发作均没有统计学差异,同时本文研究结果进一步支持低频电刺激结果中有先兆的情况与患者的预后良好显著相关,即通过电刺激诱发的先兆有利于定位致痫灶。因此,本文提出的ECoG监测中的HFOs自动检测技术和SEEG监测中发现的电刺激结果与预后的相关积极因素对于致痫灶的术前定位具有一定的临床应用价值,采用人工智能信号处理及影像-脑电融合分析等新技术将有助于更好地实现精准定位癫痫致痫灶的目标。更多还原

【Abstract】 Background:Epilepsy is a common central nervous system disease.More than 30% of epilepsy patients develop into refractory epilepsy because of ineffective drug control,and these patients often require surgical resection or amputation to induce epilepsy brain area.Therefore,accurate preoperative localization of the epileptogenic foci is the key to successful treatment of neurosurgery.In preoperative evaluation,intracranial EEG(including electrocorticogram(ECoG)and stereoelectroencephalogram(SEEG))technology plays a crucial role.For the method of ECoG,high frequency oscillations(HFOs)have been proposed as a new biomarker for locating the seizure onset zone(SOZ),which has been extensively studied in the past two decades.However,the current manual analysis method in clinic is considered to be the gold standard of HFOs detection,which is highly subjective and extremely time-consuming.Two types of signals,such as a variety of epileptiform spike waveforms,are easy to be falsely detected as HFOs,and are.For the method of SEEG,electric stimulation(ES),which is an important detection method,has not yet achieved technical unification and standardization among domestic and international medical centers,where several findings show multiple results of ES inconsistent with the relevance of surgical outcomes.Objective:In this study,a set of HFOs automatic detection algorithm based on machine learning method and its method for locating epileptogenic foci are proposed for the ECoG;meanwhile,statistical analysis on the SEEG is used to focus on the correlation between multiple factors and the prognosis of epilepsy surgery.Multiple outcomes of ES localize the association of epileptogenic foci with good prognosis.The aim of this study is to preliminarily explore the clinical application value of the automated detection of HFOs in ECoG and the statistical evaluation on the correlation between ES results and prognosis of SEEG in preoperative localization of epilepsy foci,so as to better achieve the goal of accurately locating epilepsy foci.Methods:A total of 40 patients with intracranial electrode implantation,10 ECoG patients with cortical electrode implantation and 30 SEEG patients with stereotactic electrode implantation were retrospectively included in our hospital’s neurosurgery department from September 2015 to August 2021.All patients were intractable epilepsy whose seizures could not be completely controlled by antiepileptic drug treatment and seriously affected the quality of life of the patients.Complete noninvasive evaluation methods(including video electroencephalography VEEG,MRI,CT,PET)were performed before surgery.-CT and neuropsychological examinations,etc.)still cannot be used to delineate or localize the epileptogenic zone.After shortterm,mid-term and long-term follow-up,the patients will be classified according to the Engel classification Ⅰ-Ⅳ according to the follow-up.Collect detailed clinical information and data of all patients,including seizure history,seizure symptoms,seizure frequency,seizure duration,video EEG,neuropsychological examination and related imaging examination results,and construct a set of single-center,smallsample Epilepsy patients’ intracranial cortical ECoG database and a set of intracranial stereo EEG SEEG database.In terms of ECoG,firstly,the detectedHFOs in the ECoG are preprocessed by wavelet packet filtering,and then the back propagation(BP)neural network is used to extract and automatically classify the HFOs,and then the corresponding epileptic foci are determined according to the detected high-incidence areas of HFOs;Finally,the occurrence rate of HFOs in each EEG signal channel was counted,and the obtained HFOs in the region of the EEG leads were compared with the epileptogenic foci determined by physicians through comprehensive analysis of anatomy-electro-clinical and follow-up prognosis,and evaluated the results.The detection performance of HFOs and the accuracy of epileptogenic foci location of this method.In terms of SEEG,firstly,accurate ES is based on the guidance of image modeling,and secondly,the relevant factors and prognosis are analyzed by multivariate analysis using statistical methods.The variables analyzed include age,gender,PET,scalp EEG,lateralization,and localization of the origin region,implantation site,surgical resection,pathological results,age of onset,number of electrodes,follow-up time,and high-frequency and low-frequency electrical stimulation results;finally,the statistical results were further compared in pairs,and the Bonferroni method was used to adjust theαlevel.Results:In this study,a total of 3 patients with ECoG monitoring(2 males and 1 female,age distribution 19-26 years old,average age 21.67 ± 3.09 years old)was included,which were manually marked by EEG and neurosurgeons.There are 2790 HFOs in total,and the machine learning method used in this paper can accurately detect 2511 HFOs among them,with a total of 208 false detections.The sensitivity of the obtained results is 90%,and the false detection rate is 7.46%.Among them,the EEG channel areas with high incidence of HFOs in 2 patients with focal epilepsy and the epilepsy focus area delineated by the clinician’s comprehensive assessment have a large overlap or are very close to the clinically diagnosed epilepsy focus;the other 1 patient with multifocal epilepsy has HFOs.The high-incidence channels are relatively scattered,and the estimated focus area and the overlap area with the epilepsy focus area outlined by the doctor are relatively small.Moreover,this study included 25 patients with SEEG EEG monitoring(15 males,10 females,3-37 years old,average age 22.56±8.33 years).According to the ES results,all patients were divided into three groups,including stimulation-induced seizures,auras,and the others.The prognosis of a patient is based on whether there is a seizure or not.The corresponding"yes"in the seizure or not refers to the absence of epilepsy recurrence,which means the prognosis is good(referring to the short-term and long-term follow-up being Engel grade I);the corresponding"no"in the seizure or not refers to the presence of recurrence of epilepsy represents a poor prognosis(referring to Engel classification Ⅱ-Ⅳ in the short-term and long-term follow-up).The obtained results show among the 19 patients with a good prognosis,including 14 patients with aura,4 patients with habitual seizures,and 1 patient with other patients(unstimulated aura and habitual seizures).Except for the results of low-frequency electrical stimulation,the obtained statistical results of multivariate analysis showed that there were no significant differences in other variables among groups and whether or not to have seizures(P>0.05).Moreover,the obtained results of lowfrequency electrical stimulation showed that no seizures occurred in the aura group,1 patient in the habitual seizure group(20.0%),and 5 patients in the other groups(83.3%).Fisher’s exact test showed that the differences among the three groups were statistically significant(χ~2=14.273,P<0.001).Furthermore,the results of pair-wise comparisons showed that,for adjusting theαlevel by the Bonferroni method,there was no statistical difference between the habitual seizure group and other groups,between the aura group and the habitual seizure group,and the follow-up patients in the results of low-frequency electrical stimulation(Bonferroni correction,P>0.0167);while the aura group and other groups had statistical differences with follow-up patients with seizures(Bonferroni correction,P<0.0167).Conclusion:For the method of ECoG,the preliminary research results of this study show that the machine learning could effectively detect HFOs in the cortical ECoG and could automatically locate the corresponding epileptogenic foci according to the detected high incidence area of HFOs.The epilepsy focus area(especially focal epilepsy)has a high correlation with the area determined by the clinic doctor;at the same time,the proposed method only requires the EEG monitoring of the patient during the 15-minute interval,which will substantially reduce the patients’ pain and during the long-term EEG monitoring,as well as the workload of clinical physicians to read EEG manually.For the method ES-induced SEEG,the results of the study on the correlation between multiple factors and prognosis of patients showed that,except for the results of low-frequency electrical stimulation,there were no statistical differences in other variables between groups and whether or not they had seizures.At the same time,the obtained results of this study further supporting the presence of an aura in the results of low-frequency electrical stimulation was significantly associated with good patient outcomes.That is,it would be a useful method facilitating localization of the epileptogenic foci for aura induced by electrical stimulation.Therefore,the useful automatic detection of HFOs in ECoG monitoring using machine learning and the positive factors related to electrical stimulation results and prognosis found in SEEG monitoring have certain clinical application values for preoperative localization of epileptogenic foci.New technologies such as artificial intelligence signal processing and fusion analysis of imaging-EEG would be useful in clinic to better achieve the goal of accurately locating epilepsy-causing foci.更多还原