Analysis of the microscopic images of tissues taken by biopsy is an important means for cancer diagnosis, prognosis, and grading in histopathology. Automating the analysis process completely or partially delivers significant benefits in terms of effort, time, and quantitativeness. In this talk, we introduce a fully automated method for segmentation of gland structures in microscopic tissue images. Our approach relies on representing the image as a set of primitive circular objects and use the spatio-morphological features of these primitives to determine the boundaries of the gland structures. In our method, the raw pixel data are first grouped into clusters. These clusters are then used to define a set of primitive circular objects by means of a novel tranformation technique that we introduce. Subsequently, gland candidate regions are detected by clustering the circles with respect to their spatio-morphological features. Finally, false glands are eliminated using a decision tree classifier, which is constructed on a training set. Our experiments on colon tissue images demonstrate that the proposed approach provides a more robust infrastructure for higher level image analysis, compared to its counterparts.

DATE: 14 April, 2008, Monday@ 16:40
PLACE: EA 409