Generalized N-Dimensional PCA for Compact Representation of Medical Volumes and Effective Mode Selection by Adaboost

Abstract

The paper addresses the issues of compact representation of medical volumes and effective mode selection. Medical volumes are unlike traditional two dimensional images in pattern recognition which in spatial domain has three dimensions. Meanwhile it often meets small sample problem. Thus common compact representation methods, such as PCA, do not fit medical volumes. Because the number of eigenvector is so little that it will lose much useful information. In previous work, we proposed a Generalized N-dimensional principal component analysis (GND-PCA) for reconstruction of medical volumes with only few samples. The core tensor of GND-PCA can keep most useful information. However, making diagnosis using the core tensor is difficult due to most modes very general for all samples. It will affect finally diagnosis. To resolve this problem, Adaboost is used as classifier in diagnosis because it can choose distinctive mode according to its definition. The proposed method was evaluated using a medical volume database. In our experiment, we compare Adaboost with SVM and KNN. The classification accuracy of Adaboost is slightly better than that of SVM and KNN,meanwhile time consuming of classification of Adaboost is greatly less than those of other two classification methods.