COMPUTER-AIDED DETECTION OF COLITIS ON COMPUTED TOMOGRAPHY USING A VISUAL CODEBOOK Zhuoshi Wei, Weidong Zhang, Jianfei Liu, Shijun Wang, Jianhua Yao, Ronald M. Summers Imaging Biomarkers and Computer-Aided Diagnosis Laboratory Department of Radiology and Imaging Sciences National Institutes of Health Clinical Center, USA ABSTRACT Colitis is inflammation of the colon that is frequently associated with infection and immune compromise. In this paper, we propose an automatic method for colitis detection in abdominal CT scans. We first used a visual codebook constructed by clustering feature vectors from a set of training image patches to detect the suspicious colitis regions. The initial detections included false detection points located in various organs including muscle, kidney and liver. We reduced the false positives by applying masks of these regions obtained from whole-organ segmentation. We tested our method on a CT dataset with 20 cases of colitis and 15 non-colitis cases. Average detected lesion volume for positive cases is 205ml; for negative cases is 97ml. Sixteen out of the 22 positive cases were correctly identified, yielding a sensitivity of 72.7%; 4 out of 15 negative cases were incorrectly identified, yielding a specificity of 73.3%. Index Terms— CT, colon, colitis, visual codebook 1. INTRODUCTION Abdominal computed tomography (CT) can be used to help diagnose many clinically important abnormalities including tumors, infections, injury and inflammation. Colitis is an inflammation of the colon which can have many different causes, including infection, chronic inflammation such as ulcerative colitis and Crohn's disease, immunocompromise and lack of blood flow (ischemic colitis). Colitis can be debilitating or life threatening, and early detection is essential to initiate proper treatment [1]. On CT, colitis is often associated with thickening of the colon wall. The oral contrast material trapped between nodular thickened edematous haustral folds separated by mucosal ridges is a classic radiologic sign known as the “accordion” sign [2, 3]. Fig.1 shows an example of a patient with colitis on CT. In this paper, we propose an automatic method for colitis detection in abdominal CT scans. The visual codebook has proven its effectiveness in object recognition and texture classification [4-6]. We used classifiers with a visual codebook that was trained for classifying colitis vs. different organs and tissues. The suspicious regions of colitis were first detected by the classifiers. The initial detections included false detections located in muscle, kidney and liver. We reduced the false positives by applying masks of these regions obtained from whole-organ segmentation. With a test dataset contain 20 cases of colitis and 15 non-colitis cases, our method yielded a sensitivity of 72.7% and a specificity of 73.3%. Fig.1: One example of colitis (marked by red arrows) with colon wall thickening. The oral contrast material trapped between nodular thickened edematous haustral folds separated by mucosal ridges is called the “accordion” sign by radiologists. 2. MATERIAL AND METHOD The main algorithms for colitis detection consist of two stages: initial detection for suspicious areas using a visual codebook and false detection reduction. 2.1. Visual codebook construction for colitis region The suspicious colitis areas were detected using the visual codebook framework shown in Fig.2. Given a set of training image patches, feature extraction methods were first used to obtain the image features. Since the colitis regions in our dataset have relatively coarse image textures, the size of the training image patches was selected to be 40×40 pixel grey level patches. 2013 IEEE 10th International Symposium on Biomedical Imaging: From Nano to Macro San Francisco, CA, USA, April 7-11, 2013 U.S. Government work not protected by U.S. copyright 141