基于空间和解剖结构信息的多切片核应用于阿尔茨海默症分类

Multislice kernel based on spatial and anatomical structure information for Alzheimer’s disease classification

目的 探索早期准确诊断阿尔茨海默症的快速高效算法。方法 利用脑组织的空间和解剖结构信息构造多切片核,并应用于阿尔茨海默症的分类和判别。结果 Cuingnet提出了一个框架,应用于阿尔茨海默症判别的经典单核支持向量机分类器,并将空间和解剖结构信息包含进来,生成了具有更高分类性能的可解释特征图。然而,在此框架中,为了方便使用单核模型,空间正则化参数被限制与解剖正则化参数相等。此外,高阶张量图像的矢量化破坏了固有结构,也不可避免地被迫使用一个超大规模的矩阵来定义每对体素之间的邻接关系,从而导致高额的计算复杂度。本文设计了2种新型的多切片核来改进Cuingnet框架,其中通过构造每个切片的空间和解剖拉普拉斯矩阵来保留邻接关系,并采用流行的顺序最小优化算法来估计多核学习模型参数,这样就避免了原Cuingnet框架中的超大规模矩阵计算问题。结论 实验结果表明,该改进方法将计算速度提升了数百倍,同时保持了较高的分类精度。

Objective To explore fast and efficient algorithms for early and accurate diagnosis of Alzheimer’s disease.Methods Utilizing spatial and anatomical information of brain tissue to construct multi-slice kernels and applying them to the classification and discrimination of Alzheimer’s disease.Results Cuingnet proposes a framework to include spatial and anatomical structure information in classical single kernel support vector machine for Alzheimer’s disease classification,and it generates more interpretable feature maps with high classification performances.However,in this framework,the spatial regularization parameter is restricted to be equal to the anatomical one for convenience of using single kernel model.In addition,vectorization of a higher-order tensorial image destroys the intrinsic structure and a large-scale matrix is also inevitably generated to define the adjacency relation between every pair of voxels,so it results in intensive computation loads.In this manuscript,the Cuingnet framework is improved by construction of two new types of multislice kernels wherein spatial and anatomical Laplacian matrices derived from every slice are used to retain the adjacency relations,and the widespread sequential minimal optimization algorithm is adopted to estimate the parameters in a multiple kernel learning model.In this manner,the above large-scale matrix computation in the original Cuingnet framework is avoided.Conclusion Experimental results demonstrate that computing speed is increased hundreds of times,while high classification accuracy is maintained.