|M.Sc Student||Zingman Igor|
|Subject||Analysis and Classification of Tissue Section Images|
|Department||Department of Electrical Engineering||Supervisor||Professor Ron Meir|
We consider the problem of automating the analysis and classification of microscopic biopsy images used for breast cancer diagnosis. Our classification is mainly based on visible patterns or arrangements of cell groups rather than on the morphology of individual cells. A great deal of work has been dedicated to automation in cytological analysis (dealing with information from individual cells) using computer vision methods. In contrast, the high degree of complexity of histological images prevents the widespread expansion of automated image processing tools to the analysis of tissue sections.
We present two major aspects of automated analysis: segmentation and feature extraction. In particular, we suggest two techniques for the adaptive segmentation of histological images. The first approach is based solely on intensity values and the search for an appropriate global threshold. The technique works uniformly well for a wide range of image histograms with equal and very unequal intensity modes. The second method uses a color transformation technique developed specifically to distinguish between different regions of stained tissue specimens, each with different spectral properties. This method enables the separation of bluish cellular patterns, unlike other simple methods, such as automated thresholding of hue histograms. The method is based on a model of light subtraction that predicts the unique distribution of chromacities. The prior knowledge on color distribution is used to separate the main shades of stained tissue regardless of the inconsistency of the staining strength through the specimen. The combination of intensity and color information leads to the proper segmentation of stained cellular patterns in complex histological images.
Once cellular structures are segmented, the relevant tissue features should be extracted to perform the final diagnosis. In this work we propose an original method based on the framework of mathematical morphology in order to extract features. Specifically, the concept of a morphological pattern spectrum was used to reveal features expected to contain relevant information for cancer discrimination. The pattern spectrum enables the measurement of the distribution of sizes (of a predefined pattern) within the image. We extended the regular pattern spectrum by considering the spatial information in addition to the size and shape attributes. The extended pattern spectrum yields signature maps of low dimensionality, where the information about the densities of structures of different sizes is preserved. The maps are built by an iterative procedure based on the proposed opening operator. The image signatures of cancerous tissue samples are then categorized by standard classification methods. The encouraging classification results justify the proposed features.