基于杏仁核亚区体积特征影像组学模型在癫痫分类中的应用价值研究

The Value of an Imaging Histology Model Based on Volumetric Features of the Amygdala Subregion in the Classification of Epilepsy

目的采用表面结构分割技术获得杏仁核亚区体积,对比MRI阴性颞叶癫痫(TLEneg)和海马硬化(HS)型颞叶癫痫(TLE)杏仁核亚区体积差异,构筑基于体积特征的影像组学模型来评估体积特征在癫痫分类中的价值。方法基于新疆医科大学第一附属医院癫痫数据库开展,根据纳入和排除标准筛选出受试者的临床资料以及图像,共计纳入HS颞叶癫痫(TLE-HS组)患者88例、MRI阴性颞叶癫痫(TLEneg-MRI组)患者104例、健康对照组(HC组)志愿者96名,使用FreeSurfer 7.4.1对图像进行杏仁核亚区分割处理,搜集所有亚区体积,对比三组的杏仁核亚区体积差异,最后将数据分组后导入至数坤科技平台训练和评估基于体积特征的影像组学分类模型效能。结果对比HC组,TLE-HS组和TLEneg-MRI组的部分杏仁核亚区体积显示为缩小(P<0.05),以TLEneg-MRI组体积减小更为显著;TLEneg-MRI和TLE-HS组比较,两组杏仁核亚区体积没有发现明显统计学差异,且两组杏仁核亚区改变都出现在癫痫病灶的同侧。构建各组之间的杏仁核亚区体积的影像组学模型,其中HC组和TLEneg-MRI组的效能分析中,训练集曲线下面积(AUC)为0.9596,敏感度为0.9016,特异度为0.8806,测试集AUC为0.6092,敏感度为0.6296,特异度为0.4483;而其中HS组和TLEneg-MRI组的效能分析中,训练集AUC为0.9312,敏感度为0.7869,特异度为0.9589,测试集AUC为0.7437,敏感度为0.6667,特异度为0.8065。结论不同类型的颞叶内侧癫痫能对杏仁核及其亚区体积产生不同影响,影像组学构建的模型结果提示,杏仁核的基底外侧复合体和内侧核分别有助于区分TLE伴HS和TLEneg。

Objective To obtain the amygdala subregion volume using surface structure segmentation technique,to compare the difference in amygdala subregion volume between two types of temporal lobe epilepsy,magnetic resonance nega-tive and hippocampal sclerosis,and to construct an image histological model based on volume features to evaluate the value of volume features in epilepsy classification.Methods Conducted based on the epilepsy database of the First Affiliated Hospital of Xinjiang Medical University,the clinical data as well as images of subjects were screened according to the inclusion and exclusion criteria,and a total of 88 patients with hippocampal sclerosis-type temporal lobe epilepsy(TLE-HS group),104 patients with magnetic resonance-negative temporal lobe epilepsy(TLEneg-MRI group),and 96 volunteers from the healthy control group(HC group)were enrolled,using the FreeSurfer 7.4.1 to segment the amygdala subregions of the images,collect all the subregion volumes,compare the amygdala subregion volume differences among the three groups,and finally group the data and import them into the DigitalKun Technology platform to train and evaluate the efficacy of the volumetric feature-based classification model for imaging genomics.Results Compared with the HC group,some of the amygdala subregion volumes of the HS group and TLE-MRIneg group showed a reduction(P<0.05),and the volume reduction was more significant in the TLE-MRIneg group.No statistically significant difference was found in the amygdala subregion volumes of the two groups when comparing the TLE-MRIneg and HS groups,and the changes in the amygdala subregions of the two groups were found to be on the ipsilateral side of the epileptic lesions.Imaging histological models of amygdala subregion volume were constructed between the groups,in which the HC group and the TLE-MRIneg group had a training set AUC of 0.9596,a sensitivity of 0.9016,and a specificity of 0.8806,and the test set had an AUC of 0.6092,a sensitivity of 0.6296,and a specificity of 0.4483 in the efficacy analysis,whereas the HS group and the TLE-MRIneg group set AUC was 0.9312,sensitivity was 0.7869,and specificity was 0.9589,while the test set AUC was 0.7437,sensitivity was 0.6667,and specificity was 0.8065.Conclusion Different types of medial temporal lobe epilepsy can differentially affect the volume of the amygdala and its subregions,and the results of the model constructed by imaging histology suggest that the basolateral complex and the medial nucleus of the amygdala are useful in distinguishing temporal lobe epilepsy with hippocampal sclerosis from negative temporal lobe epilepsy,respectively.