Medical Image Segmentation using Cooccurrence Matrix based Texture Features Calculated on Weighted Region

L. Tesar (Japan, Czech Republic), D. Smutek (Czech Republic), A. Shimizu, and H. Kobatake (Japan)


bioinformatics, medical image analysis, cooccurrence ma trix, image segmentation, texture features, Haralick fea tures, 3D image analysis


An improvement of texture based 2D or 3D image segmen tation method is proposed, aimed at medical image analy sis. Segmentation of organs or disease diagnosing is the tar getted problem. Features based on cooccurrence matrix are studied. A new approach to local cooccurrence matrix cal culation is proposed in this paper. Separate cooccurrence matrix is calculated for each pixel (in 2D case) or vowel (in 3D case), based on pixels (voxels) around the original pixel (voxel). The new approach is in the way of calculat ing the cooccurrence matrix. As opposed to classical defi nition where cooccurrence matrix values are counts of rel ative frequencies of cooccurrences over defined region, we propose for each cooccurrence, to be weighted by the func tion of distance from the original pixel (voxel) of interest. Compared to calculating the cooccurrence texture features over square (or cubic) region around a pixel (voxel), the proposed approach makes cooccurrence matrix and texture feature more focused, therefore it can be used to search for smaller regions with different texture properties (like tumours). A set of abdominal CT images is used for evalu ation of the proposed approach and comparison with older cubic region based approach. The actual cooccurrence fea tures used were some of those proposed by Haralick.

Important Links:

Go Back