Computerized Renal Cell Carcinoma Nuclear Grading Using 3D Textural Features Tae Yun Kim, and Heung Kook Choi Department of Computer Science, Inje University, Gimhae, Korea Ubiquitous Healthcare Research Center, Inje University, Gimhae, Korea Abstract— An extraction of important features in cancer cell image analysis is a key process in grading renal cell carcinoma. In this study, we applied three-dimensional (3D) texture feature extraction methods to cancer cell nuclei images and evaluated the validity of them for computerized cell nuclei grading. Individual images of 1,800 cell nuclei were extracted from 8 classes of renal cell carcinomas (RCCs) tissues using confocal laser scanning microscopy (CLSM). First, we extracted the chromatin texture quantitatively by calculating 3D gray-level co-occurrence matrices (3D GLCM) and 3D run length matrices (3D GLRLM). To demonstrate the suitability of 3D texture features for grading, we had performed a principal component analysis to reduce feature dimensionality, then, we also performed discriminant analysis as statistical classifier. Finally this result was compared with the result of classification using several optimized features that extracted from stepwise features selection. Additionally AUC (area under curve) analysis was performed for the grade 2 and 3 cell images. Three dimensional texture features have potential for use as fundamental elements in developing a new nuclear grading system with accurate diagnosis and predicting prognosis. Keywords- Digital Image Cytometry, 3D Texture Features, Statistical Analaysis I. INTRODUCTION The evaluation of cell nuclei features helps determine the prognoses of patients with carcinomas, and from a cytometric perspective, the variation in the distribution of chromatin is considered a very important characteristic. Various computer-assisted diagnosis systems have been developed, and in most of these systems, feature extraction is an important means of classification because the classification performance depends on the extracted features. The most critical problems with the available systems are that they still depend for the most part on the pathologist's subjective decision. Conventional visual analysis for grading has low reproducibility because it is based on subjective evaluation, which is prone to inter- and intra- observer variation, regardless of any grading system. Several studies have examined and defined various features of cell nuclei such as the internal structure of cells (granularity and regularity of chromatin), size irregularity, shape of the nucleus, and distance between nuclei, which are important for determining the progress of cancer [1-7]. Histology-based statistical analyses of textural features are generally based on the gray level of cell nuclei, while the structural analysis method describes the properties and placement of texture elements. Nevertheless, most two- dimensional (2D) texture feature based analysis systems still have low objectivity and reproducibility. Given the variety of analysis methods, no clear measurement standard has been established for extracting accurate numerical information. Similarly, the image analysis systems based on 2D images have several intrinsic limitations. For example, cells and cell nuclei are not perfectly spherical, and consequently, their shape differs noticeably according to the cutting angle and thickness of the sample tissues. Ultimately, it is difficult to confirm the shape of a cell. Another drawback of conventional 2D slice-based approaches is that they are tedious, fatiguing, and time-consuming. To guarantee reproducibility, a new method based on three- dimensional (3D) image analysis is required. Recently, some papers have reported different 3D texture features applied to various medical images. Jafari- Khouzani et al. suggested an analysis method based on a comparative study of 2D and 3D wavelet features [8]. Madhabushi et al. studied the automatic segmentation of high-resolution magnetic resonance (MR) images using a 3D Gabor filter and a co-occurrence matrix [9]. Kurani and Xu applied a 3D gray-level co-occurrence matrix (GLCM) and a 3D gray-level run length method (GLRLM) to computed tomography (CT) images to separate various organs of the human body [10-11]. Most of these approaches were simply extended from conventional 2D methods, but the importance of 3D texture increases with its successful expansion. This study evaluates the validity of 3D GLCM and 3D GLRLM, which were studied by Kurani and Xu, by applying them to the images of cell nuclei for RCC (Renal Cell Carcinoma) obtained by CLSM. Our previous study examined the correlation between the changing grade following the cancer process and the 3D morphological features and also investigated the 2D features that can be a good proxy for estimating 3D features, while mainly focusing on the morphological changes of the This work was supported by the Korea Research Foundation Grant funded by the Korean Government (MOEHRD, Basic Research Promotion Fund) (KRF-2006-311-D00840) Tae Yun Kim is Ph.D. student at the department of computer science, Inje University, Gimhae, Korea (e-mail: liminus@paran.com) Heung Kook Choi is associate professor at the department of computer science, Inje University, Gimhae, Korea (e-mail: cschk@inje.ac.kr) U.S. Government work not protected by U.S. copyright U.S. Government work not protected by U.S. copyright