Graphics Reference
In-Depth Information
slide is usually much higher than common digital image and the characteristics of pa-
thology image are much different from natural images.
To deal with the retrieval problem on digital pathology slide databases, Content-
Based Image Retrieval (CBIR) has been proposed and successfully applied to clinical
diagnosis [4, 5]. Over the past years, a large number of retrieval methods for pathology
image have been proposed. Caicedo et al. [6] apply different kinds of visual features to
achieve the retrieval task for four kinds of tissues. Kowal et al. [7] take advantage of
statistical features of individual nuclei to classify benign and malignant cases of breast
cancer. However, these methods mentioned above just describe the global characteris-
tics of the image and may even ignore the high-level semantics that exist in the image.
Therefore, to mine the texture information and local property of pathology image, we
propose to divide the entire image into non-overlapping blocks and extract Local Bi-
nary Pattern (LBP) [8] descriptor in each block. Then LBP descriptors are used to build
the codebook composed of visual words through k-means. Afterwards, each pathology
image can be represented by the word frequency histogram via Bag-of-Words (BoW)
[12] scheme. However, there are synonyms among visual words and thus make the
word-level representation hard to discriminatively reveal the semantics in images.
Therefore, probabilistic latent semantic analysis (pLSA) [9] model is applied in our
method to mine the high-level semantics of words. Nonetheless, pLSA model just uses
BoW scheme to discover the word distribution, which is likely to ignore that there are
some correlated and specific characteristics among words. Consequently, the word-
level representation of conventional pLSA model may fail to describe the image con-
tent precisely. To improve the discriminant ability of pLSA, we apply low-rank and
sparse matrix decomposition technique [10, 13] to decompose the word-level represen-
tation into two meaningful parts (i.e., correlated and specific word-level representa-
tions), and then utilize them to train two pLSA models. Finally each image can be
represented by the combination topics learned from these two models.
Our proposed method consists of two contributions. First, we use block LBP fea-
tures to describe the spatial texture information and then apply pLSA model to dis-
cover the high-level semantics of pathology images. The second is that we use the
low-rank and sparse matrix decomposition to obtain two word-level representations
which can characterize the correlated and specific parts of the visual word distribu-
tion. As a consequence, the discriminant ability of word-level representation has been
greatly improved. Experimental results on the digital pathology image database of
breast cancer demonstrate the feasibility and effectiveness of our method.
The rest of the paper is arranged as follows: Section 2 introduces block LBP de-
scriptor. Section 3 describes the pLSA model along with the low-rank and sparse
matrix decomposition. Section 4 presents the experimental database and results. Fi-
nally the conclusion is given in Section 5.
2
Block Local Binary Pattern (LBP)
Local binary Pattern (LBP) [8] is a powerful local texture descriptor with the advan-
tages of rotation invariance and orientation invariance. The pattern of each pixel is
Search WWH ::
Custom Search