Image Display, Enhancement, and Analysis (IDEA) Laboratory, Biomedical Research Imaging Center (BRIC) and Department of Radiology, University of North Carolina at Chapel Hill, NC, USA.
Neuroimage. 2012 Feb 1;59(3):2045-56. doi: 10.1016/j.neuroimage.2011.10.015. Epub 2011 Oct 14.
Different imaging modalities provide essential complementary information that can be used to enhance our understanding of brain disorders. This study focuses on integrating multiple imaging modalities to identify individuals at risk for mild cognitive impairment (MCI). MCI, often an early stage of Alzheimer's disease (AD), is difficult to diagnose due to its very mild or insignificant symptoms of cognitive impairment. Recent emergence of brain network analysis has made characterization of neurological disorders at a whole-brain connectivity level possible, thus providing new avenues for brain diseases classification. Employing multiple-kernel Support Vector Machines (SVMs), we attempt to integrate information from diffusion tensor imaging (DTI) and resting-state functional magnetic resonance imaging (rs-fMRI) for improving classification performance. Our results indicate that the multimodality classification approach yields statistically significant improvement in accuracy over using each modality independently. The classification accuracy obtained by the proposed method is 96.3%, which is an increase of at least 7.4% from the single modality-based methods and the direct data fusion method. A cross-validation estimation of the generalization performance gives an area of 0.953 under the receiver operating characteristic (ROC) curve, indicating excellent diagnostic power. The multimodality classification approach hence allows more accurate early detection of brain abnormalities with greater sensitivity.
不同的成像方式提供了重要的互补信息，可用于增强我们对大脑疾病的理解。本研究专注于整合多种成像方式，以识别有轻度认知障碍（MCI）风险的个体。MCI 通常是阿尔茨海默病（AD）的早期阶段，由于其认知障碍的症状非常轻微或不明显，因此难以诊断。最近出现的脑网络分析使得在全脑连接水平上对神经疾病进行特征描述成为可能，从而为脑疾病分类提供了新途径。我们使用多核支持向量机（SVM）尝试整合扩散张量成像（DTI）和静息状态功能磁共振成像（rs-fMRI）的信息，以提高分类性能。结果表明，与独立使用每种模态相比，多模态分类方法在准确性方面具有统计学上的显著提高。所提出方法的分类准确率为 96.3%，比基于单模态的方法和直接数据融合方法至少提高了 7.4%。接收器操作特性（ROC）曲线下的交叉验证估计给出了 0.953 的区域，表明具有出色的诊断能力。因此，多模态分类方法可以提高敏感性，更准确地早期检测大脑异常。
