Abstract:  Machine learning has been widely applied to the diagnosis of brain diseases at present. The application of medical imaging analysis with limited samples and large feature dimension is a typical small sample problem. In addition, multi-modal data (such as MRI and PET) can be obtained via different imaging procedures. This paper proposes a novel feature selection method to handle the small sample of multi-modal imaging data. Firstly, each modality is treated as a single learning task and the l_2,1 norm is used to jointly select features, which can guarantee the features of different modalities in the same brain regions to be selected at the same time. And hypergraph learning is involved to capture the high-order relationships among the samples, which can depict the priori distribution of data in each modality. Then, the multi-kernel support vector machine (SVM) is used to fuse the features chosen from multi-modal data for final classification. The proposed method is validated and compared with traditional algorithms on ADNI dataset, and the experimental results show the efficiency of the proposed method.

Key words: hypergraph learning, multi-task learning, feature selection, multi-modal classification, Alzheimer’s disease

摘要： 目前机器学习算法已经被广泛应用到脑疾病的诊断中。医学影像数据由于样本珍贵，并且特征维数往往远大于已有样本数目，在实际应用中这是典型的小样本问题。此外，通过不同的成像手段可以得到不同模态的数据（例如MRI和PET）。从而提出一种基于超图的多模态特征选择算法。首先将每组模态当作一组任务，利用 l_2,1 范数进行特征选择，保证不同模态相同脑区的特征被选中。然后使用超图技术来刻画数据样本与样本之间的高阶信息，从而充分利用每组模态数据内部的分布先验。最后利用多核支持向量机对选择后的特征进行融合分类，从而提高对疾病的诊断精度。在ADNI数据集上对提出的方法进行验证，并与传统方法进行对比，实验结果说明了提出方法的有效性。

关键词: 超图学习, 多任务学习, 特征选择, 多模态分类, 阿尔茨海默症