Appearance Models for Medical Volumes with Few Samples by Generalized 3D-PCA

Appearance models is important for the task of medical image analysis, such as segmentation. Principal component analysis (PCA) is an efficient method to build the appearance models; however the 3D medical volumes should be first unfolded to form the 1D long vectors before the PCA is used. For large medical volumes, such a unfolding preprocessing causes two problems. One is the huge burden of computing cost and the other is bad performance on generalization. A method named as generalized 3D-PCA is proposed to build the appearance models for medical volumes in this paper. In our method, the volumes are directly treated as the third order tensor in the building of the model without the unfolding preprocessing. The output of our method is three matrices whose columns are formed by the orthogonal bases in the three mode subspaces. With the help of these matrices, the bases in the third order tensor space can be constructed. According to these properties, our method is not suffered from the two problems of the PCA-based methods. Eighteen 256×256×26 MR brain volumes are used in the experiments of building appearance models. The leave-one-out testing shows that our method has good performance in building the appearance models for medical volumes even when few samples are used for training.