TEXTURE CLASSIFICATION USING NONLINEAR COLOR QUANTIZATION: APPLICATION TO HISTOPATHOLOGICAL IMAGE ANALYSIS Olcay Sertel 1,2 , Jun Kong 1,2 , Gerard Lozanski 3 , Arwa Shana’ah 3 , Umit Catalyurek 1,2 , Joel Saltz 2 , Metin Gurcan 2 1 Department of Electrical and Computer Engineering 2 Department of Biomedical Informatics 3 Department of Pathology, The Ohio State University, Columbus, OH 43210, USA ABSTRACT In this paper, a novel color texture classification approach is introduced and applied to computer-assisted grading of follicular lymphoma from whole-slide tissue samples. The digitized tissue samples of follicular lymphoma were classified into histological grades under a statistical framework. The proposed method classifies the image either into low or high grades based on the amount of cytological components. To further discriminate the lower grades into low and mid grades, we proposed a novel color texture analysis approach. This approach modifies the gray level co- occurrence matrix method by using a non-linear color quantization with self-organizing feature maps (SOFMs). This is particularly useful for the analysis of H&E stained pathological images whose dynamic color range is considerably limited. Experimental results on real follicular lymphoma images demonstrate that the proposed approach outperforms the gray level based texture analysis. Index Terms— color texture analysis, self-organizing feature maps, computer-aided diagnosis 1. INTRODUCTION Follicular Lymphoma (FL) is a one of the most common non-Hodgkin B cell lymphomas in the western world with a highly variable clinical course. Patients with indolent FL do not benefit from early therapy. In fact, early chemotherapy for them may cause more harms than benefits; therefore should be avoided. On the other hand, FL patients with aggressive disease should receive appropriate therapy as soon as possible to increase their chance of remission and to prolong their lives. These important clinical decisions are currently guided by histological grading of the tumor. As recommended by the World Health Organization (WHO), histological grading of FL is based on the number of large malignant cells, namely centroblasts (CB), per standard 40 high power microscopic field (HPF) of 0.159 mm 2 . In this method, centroblasts are manually counted in ten random neoplastic follicles and the average of CB/HPF [1] is reported. In this grading system, grade I corresponds to 5 or less CB/HPF, grade II to 6-15 CB/HPF and grade III to 15 or more CB/HPF. Although it is very important in clinical practice, this manual method suffers from well-documented inter- and intra-reader variability. For instance, in a multi- site study, the agreement among experts for the various grades of follicular lymphoma varied between 61% and 73% [2]. Moreover, for practical reasons, pathologists typically count CBs only in ten neoplastic follicles, leading to sampling bias. Possible consequences of over or under grading of FL include inappropriate timing and type of therapy with serious clinical consequences for patients. Therefore, we are developing a computer-assisted system for automated grading of FL with a better consistency. Parallel to the developments in digital scanning technologies, research on histopathological image analysis is becoming more and more active. Recently, several image analysis approaches have been proposed for different types of cancers such as prostate [3], neuroblastoma [4] and colon cancers [5]. All of these studies exploit the texture information and construct the subsequent analysis over a statistical classification framework. However, most of the texture models are derived from gray-level images. The color information is incorporated after separating the color from the illumination from which the texture information is extracted and combined with the color information. Among many texture models, gray-level co-occurrence method introduced by Haralick et al. is one of the most widely used texture analysis approach [6]. However, this approach is limited to gray-level images. Arvis et al. proposed a multi-spectral method and a uniform quantization method to incorporate the color and the texture information in the co-occurrence matrix framework [7]. The basic idea behind the multi-spectral method is to use the cross-correlation between channels to construct several co- occurrence matrices. In the latter approach, instead of using the gray-levels, color images are quantized to extract several color classes and the co-occurrence matrix uses the label of the classes for its computation. These studies conclude that the color texture approach improves the performance remarkably. In this paper, we propose to use the color texture information from H&E-stained images for the automated 597 1-4244-1484-9/08/$25.00 ©2008 IEEE ICASSP 2008