Bilkent University
Department of Computer Engineering


Dominant Blob Scale Histograms for Tissue Image Classification


Tunç Gültekin
MSc Student
Computer Engineering Department
Bilkent University

Different grades of cancer cause shape and size deformations on tissue components at different levels. Correct characterization of these structural deformations is important for accurate cancer diagnosis and grading. In this work, first, we introduce a new set of feature descriptors for modeling shape and size of tissue components to classify tissue images. Our new descriptor set finds the scales of objects by fitting difference of Gaussians, which is called Mexican Hat Shapes, on to the tissue components. Second, we define our new feature descriptor set on different Gabor responses of tissue components and quantify the frequency of two component types being co-occurred in Delaunay triangulation relationship. Finally, we combine this information into a single feature vector to classify six different categories of cancer. Our experiments reveal that our definition of these new descriptors provides better representation for tissue images. This is attributed to the property of Gaussians better quantifying different shapes and scales of tissue components and more effectively removing pixel-based noise in tissue images. Our experiments also show the use of scale and shape modeling through the proposed descriptors leads to better classification results especially in cancer categorization


DATE: 04 November, 2013, Monday @ 15:40