Automated detection of hippocampal sclerosis using clinically empirical and radiomics features
Automated detection of hippocampal sclerosis using clinically empirical and radiomics features
复制标题
使用临床经验和放射组学特征自动检测海马硬化。
DOI：
10.1111/epi.16392
复制
发表时间：
2019-12-01
期刊：
影响因子：
5.6
通讯作者：
Tian, Jie
中科院分区：
文献类型：
--
作者：
Mo, Jiajie;Liu, Zhenyu;Tian, Jie
目的：颞叶癫痫是一种常见的癫痫类型，可能适合手术治疗。然而，磁共振成像（MRI）阴性的海马硬化症（HS）在临床实践中会妨碍患者的早期诊断和手术干预，导致疾病进展。我们的目标是自动检测和评估HS的结构改变。方法：80例经病理证实的难治性癫痫患者和80例健康对照者纳入研究。两个自动分类器依赖于临床经验和放射组学特征的开发，以检测HS。对所有参与者进行交叉验证，并在80个对照组中评估特异性。还评价了模型的性能、稳健性和临床实用性。进行结构分析以研究HS的形态学异常。结果如下：基于临床经验特征的计算模型表现出优异的性能，主要队列的曲线下面积（AUC）为0.981，验证队列为0.993。其中一个特征，即颞极的灰白质边界模糊，在模型性能中表现出最高的权重。另一个基于放射组学特征的模型也显示出令人满意的性能，主要队列中的AUC为0.997，验证队列中的AUC为0.978。特别是，该模型将MRI阴性HS的检出率提高到96.0%。颞极皮层折叠复杂性的新特征不仅在分类器中起着至关重要的作用，而且与疾病持续时间有显着的相关性。意义：具有定量临床和放射组学特征的机器学习可改善HS检测。HS相关的结构改变在MRI阳性和MRI阴性HS患者组中相似，表明误诊主要源于经验解释。颞极皮质折叠的复杂性是一个潜在的有价值的功能，探索HS的性质。
Objective: Temporal lobe epilepsy is a common form of epilepsy that might be amenable to surgery. However, magnetic resonance imaging (MRI)-negative hippocampal sclerosis (HS) can hamper early diagnosis and surgical intervention for patients in clinical practice, resulting in disease progression. Our aim was to automatically detect and evaluate the structural alterations of HS. Methods: Eighty patients with pharmacoresistant epilepsy and histologically proven HS and 80 healthy controls were included in the study. Two automated classifiers relying on clinically empirical and radiomics features were developed to detect HS. Cross-validation was implemented on all participants, and specificity was assessed in the 80 controls. The performance, robustness, and clinical utility of the model were also evaluated. Structural analysis was performed to investigate the morphological abnormalities of HS. Results: The computational model based on clinical empirical features showed excellent performance, with an area under the curve (AUC) of 0.981 in the primary cohort and 0.993 in the validation cohort. One of the features, gray-white matter boundary blurring in the temporal pole, exhibited the highest weight in model performance. Another model based on radiomics features also showed satisfactory performance, with AUC of 0.997 in the primary cohort and 0.978 in the validation cohort. In particular, the model improved the detection rate of MRI-negative HS to 96.0%. The novel feature of cortical folding complexity of the temporal pole not only played a crucial role in the classifier but also had significant correlation with disease duration. Significance: Machine learning with quantitative clinical and radiomics features is shown to improve HS detection. HS-related structural alterations were similar in the MRI-positive and MRI-negative HS patient groups, indicating that misdiagnosis originates mainly from empirical interpretation. The cortical folding complexity of the temporal pole is a potentially valuable feature for exploring the nature of HS.