Semantic Feature Extraction for Brain CT Image Clustering Using Nonnegative Matrix Factorization
Brain computed tomography (CT) image based computer-aided diagnosis (CAD) system is helpful for clinical diagnosis and treatment. However it is challenging to extract significant features for analysis because CT images come from different people and CT op
- PDF / 1,727,374 Bytes
- 8 Pages / 430 x 659.996 pts Page_size
- 87 Downloads / 182 Views
Research Center of Biomedical Engineering, Graduate School at Shenzhen Tsinghua University, Shenzhen, China 518055 [email protected] 2 mTools Ltd., Suite 2418, No.102, Xian Lie Middle Rd. Guangzhou, China 510070
Abstract. Brain computed tomography (CT) image based computeraided diagnosis (CAD) system is helpful for clinical diagnosis and treatment. However it is challenging to extract significant features for analysis because CT images come from different people and CT operator. In this study, we apply nonnegative matrix factorization to extract both appearance and histogram based semantic features of images for clustering analysis as test. Our experimental results on normal and tumor CT images demonstrate that NMF can discover local features for both visual content and histogram based semantics, and the clustering results show that the semantic image features are superior to low level visual features.
1
Introduction
Brain computed tomography (CT) image based computer-aided diagnosis (CAD) system is helpful for clinical diagnosis and treatment. Many brain CT CAD systems depend on content based image retrieval (CBIR) system [1,2]. In CBIR system, images are always processed and analyzed with visual features, including original gray/color image, color space, texture, shape feature, regions, spatial relation features, even volume of interest, and their combination or fusion [3,4,5,6]. In recent years, semantic based image analysis has been proposed to close the gap between low level visual content and high level semantics of an image [4,7,8]. In representation of an image, the original and low level visual features always have large size. For example, for a 256 × 256 gray image, it contains 65536 pixel values. On the contrast, high level semantic features are more compact, and always less than such size. For example, histogram with several hundred scale levels (e.g. 256 scales or bins), is one approach for automatically extracting low level semantic features from visual features. For example, in [9] color histograms are adopted for image classification. And in [10], latent semantic indexing (LSI)
Corresponding author.
D. Zhang (Ed.): ICMB 2008, LNCS 4901, pp. 41–48, 2007. c Springer-Verlag Berlin Heidelberg 2007
42
W. Liu et al.
is used for semantic feature extraction on color histogram based feature-imagematrix. However LSI is essentially based on single value decomposition (SVD) or principal component analysis (PCA), and it contains negativity values in decomposition, missing physical meanings. Recently nonnegative matrix factorization (NMF) has been proven a powerful method for nonnegative data, such as images and documents [11]. NMF has parts based representation which can find local features with contrast to PCA. In addition, NMF is similar to K-means clustering method and essentially a soft clustering approach [11,12]. In a recent work [13], NMF is used for feature extraction on both visual content and histogram in image retrieval. However, how to find local semantic features from such color his
Data Loading...
