Classification of skin tumors based on shape features of nuclei

In this paper, an automated system to segment cell nuclei of dermatofibroma (DF) and dermatofibrosarcoma protuberans (DFSP) is proposed. It is difficult to segment nuclear regions accurately, because there exists a lot of ambiguous nuclei. Objective of the system is to segment nuclear regions objectively and in some accuracy. In the system, regions to be segmented are defined as regions that are surrounded by edges of certain strength. Under this restriction, arbitrary shaped nuclear regions and weakly stained nuclear regions are possible to be extracted. At first, contrast emphasis using hue is done as preprocessing. Image binarization is done by a dynamic thresholding method with implementation of Laplacian-histogram method and Otsu□fs thresholding method. At last, Separation of the overlapping nuclei is carried out by watershed algorithm. To evaluate availability of this system, segmentation test was done using real tissue cell images of DF and DFSP. From some features computed from the segmented nuclear regions, an automated classification of benign tumor or malignant tumor is finally performed.