Tensor-Based Subspace Learning for Classification of Focal Liver Lesions in Multi-phase CT Images

Jian Song,Sihang Zhu,Lanfen Lin,Hongjie Hu,Yen-Wei Chen

DOI: https://doi.org/10.1007/978-3-030-32456-8_66

2019-01-01

Abstract:Medical images play an important role in clinics. Machine learning has been widely used in the fields of computer vision and pattern recognition and computer-aided diagnosis with medical images become an active research topic. Efficient representation of the medical images or effective extraction of discriminative features from CT images is one of crucial steps in computer-aided diagnosis. Principal component analysis (PCA) is a subspace learning method and is widely used for efficient representation of data. The limitation of PCA is that a multi-dimensional data (e.g. an image or a video image) should be unfolded into a vector resulting in loss of spatial and spatial-temporal relationship of the data. In this paper, we proposed an efficient representation of multi-phase CT images based on a tensor-based subspace learning method known as generalized N-dimensional principal component analysis (GND-PCA). In the proposed method, the multi-phase CT image is treated a tensor without vector-unfolding for subspace learning. The core tensor obtained by GND-PCA is used as temporal and spatial features for focal liver lesion classification. Experiments show that in the case of fewer samples, GND-PCA achieved better results than conventional PCA and 2D-PCA, which is an extension of PCA.