Morphological Hippocampal Markers for Automated Detection of Alzheimer's Disease and Mild Cognitive Impairment Converters in Magnetic Resonance Images
Morphological Hippocampal Markers for Automated Detection of Alzheimer's Disease and Mild Cognitive Impairment Converters in Magnetic Resonance Images
复制标题
DOI：
10.3233/jad-2009-1082
复制
发表时间：
2009-07-01
影响因子：
4
通讯作者：
Milles, Julien
中科院分区：
文献类型：
--
作者：
Ferrarini, Luca;Frisoni, Giovanni B.;Milles, Julien
在这项研究中，我们研究了使用海马形状为基础的标志物自动检测阿尔茨海默病（AD）和轻度认知障碍转换（MCI-c）。采集50例AD受试者、50例年龄匹配对照、15例MCI-c和15例MCI-nc的三维t1加权磁共振图像。从归一化图像中获得两个海马的手动描绘。全自动形状建模用于生成两个结构的可比网格。在随机抽样的训练集（25个对照组和25个ad）上进行重复排列测试，突出显示了基于形状的标记，这些标记主要位于CA1扇区，可以始终区分ad和对照。支持向量机（svm）训练，使用标记从一个或两个海马体，自动分类控制和AD受试者。对其余25个ad和25个对照进行留1交叉验证，左侧海马体标记的最佳准确性为90%（灵敏度为92%）。使用相同的形态学标记来训练支持向量机进行MCI-c和MCI-nc分类：右侧海马体中的标记达到80%的准确性（和灵敏度）。由于模式识别框架，我们的结果在统计上代表了临床设置的预期性能，并且与基于海马体积的分析相比较有利。
In this study, we investigated the use of hippocampal shape-based markers for automatic detection of Alzheimer's disease (AD) and mild cognitive impairment converters (MCI-c). Three-dimensional T1-weighted magnetic resonance images of 50 AD subjects, 50 age-matched controls, 15 MCI-c, and 15 MCI-non-converters (MCI-nc) were taken. Manual delineations of both hippocampi were obtained from normalized images. Fully automatic shape modeling was used to generate comparable meshes for both structures. Repeated permutation tests, run over a randomly sub-sampled training set (25 controls and 25 ADs), highlighted shape-based markers, mostly located in the CA1 sector, which consistently discriminated ADs and controls. Support vector machines (SVMs) were trained, using markers from either one or both hippocampi, to automatically classify control and AD subjects. Leave-1-out cross-validations over the remaining 25 ADs and 25 controls resulted in an optimal accuracy of 90% (sensitivity 92%), for markers in the left hippocampus. The same morphological markers were used to train SVMs for MCI-c versus MCI-nc classification: markers in the right hippocampus reached an accuracy (and sensitivity) of 80%. Due to the pattern recognition framework, our results statistically represent the expected performances of clinical set-ups, and compare favorably to analyses based on hippocampal volumes.