Automated Retrieval of Focal Liver Lesions in Multi-phase CT Images Using Tensor Sparse Representation

Content based image retrieval (CBIR) that searches for similar images in a large database has been attracting increasing research interest recently, and it has been applied to medical image characterization for sharing experts’ experiences. One challenging task in CBIR is to extract features for effective image representation. To this end, bag-of-visual-words (BoVW) has been proven to be effective to extract middle-level features for image analysis. However, it is necessary to first vectorize the two- or three-dimensional spatial structure for analysis in conventional BoVW and then destroy the spatial relationships of nearby voxels. In this study, we propose a tensor sparse coding method, which is a multilinear generalization of conventional sparse coding (soft assignment in BoVW), to learn features from multi-dimensional medical images. We regard high-dimensional local structures as tensors and propose a K-CP (CANDECOMP/PARAFAC) algorithm to learn an overcomplete tensor dictionary iteratively. By using the learned overcomplete tensor dictionary, sparse coefficients of tensor local structures are calculated by employing the tensor orthogonal matching pursuit (Tensor-OMP) algorithm, which is an extended multilinear version of the conventional vector-based OMP. The proposed method is applied to the retrieval of focal liver lesions (FLLs) by using a medical database consisting of contrast-enhanced multi-phase computer-tomography (CT) images. Experiments show that the proposed tensor sparse coding method achieved better retrieval performance than conventional methods.