在利用图像进行疾病针对的研究中，不同的研究选择利用了不同的图像模态，不同的特征提取和选择方法，不同的分类器进行试验。针对现有的方法集中在对不同模态的图像在特征水平上进行拼接和提高所用特征的多样性上，而实际上不同的分类器也对分类效果提供了互补的信息的观察。我们设计了一种基于元学习的方法在分类器的水平上进行分类。通过结合基于装袋堆叠（Bagging of Stacking）的投射算法和K近邻（KNN），我们的方法同时考虑了图像模态，特征和分类器的多样性。更重要的是，由于我们的方法实现了方法和底层特征之间的分离，现有的各种方法都可以扩展到我们的分类框架中。基于老年痴呆病人图像分类试验展示了我们算法的有效性。
英文摘要
For medical image classification, the classification methods have been built on a variety of features extracted from data of multiple imaging modalities with different classification techniques. we proposed a meta-learning strategy for fusing multiple AD classifiers, referred to as base classifiers, which might be built on image features in different representations from data of diverse imaging modalities with various classification techniques. Instead of directly combining base classifiers, we extract discriminative features from images by mapping image features in different representations adopted in base classifiers with bagging-of-stacking projections (BSPs). Each of BSPs is essentially an overfitting-resistant stacking classifier of base classifiers built on bagging training samples. Built on the discriminative features extracted by BSPs, a k-nearest neighbor (KNN) classifier has demonstrated promising classification performance in AD classification.
修改评论