Image Dissimilarity-Based Quantification of Lung Disease from CT Lauge Sørensen 1 , Marco Loog 1,2 , Pechin Lo 1 , Haseem Ashraf 3 , Asger Dirksen 3 , Robert P.W. Duin 2 , and Marleen de Bruijne 1,4 1 The Image Group, Department of Computer Science, University of Copenhagen, Denmark lauges@diku.dk 2 Pattern Recognition Laboratory, Delft University of Technology, The Netherlands 3 Department of Respiratory Medicine, Gentofte University Hospital, Denmark 4 Biomedical Imaging Group Rotterdam, Departments of Radiology & Medical Informatics, Erasmus MC, The Netherlands Abstract. In this paper, we propose to classify medical images using dissimilarities computed between collections of regions of interest. The images are mapped into a dissimilarity space using an image dissimilarity measure, and a standard vector space-based classifier is applied in this space. The classification output of this approach can be used in com- puter aided-diagnosis problems where the goal is to detect the presence of abnormal regions or to quantify the extent or severity of abnormalities in these regions. The proposed approach is applied to quantify chronic obstructive pulmonary disease in computed tomography (CT) images, achieving an area under the receiver operating characteristic curve of 0.817. This is significantly better compared to combining individual re- gion classifications into an overall image classification, and compared to common computerized quantitative measures in pulmonary CT. 1 Introduction Quantification of abnormality in medical images often involves classification of regions of interest (ROIs), and combination of individual ROI classification out- puts into one global measure of disease for the entire image [1,2,3,4,5,6,7]. These measures may, e.g., express a probability of the presence of certain abnormalities or reflect the extent or severity of disease. A global image measure based on the fusion of several independent ROI clas- sifications disregards the fact that the ROIs belong to a certain image in the classification step. Moreover, in some cases only global image labels are avail- able, while the images are still represented by ROIs in order to capture localized abnormalities. In some studies, this is handled by propagating the image label to the ROIs within that image, which again allows fusion of individual ROI clas- sifications, to obtain a global image measure [4,5,6]. However, an image showing abnormality will generally comprise both healthy and abnormal regions, and the above approach, incorrectly, labels ROIs without abnormality in such an image as abnormal. T. Jiang et al. (Eds.): MICCAI 2010, Part I, LNCS 6361, pp. 37–44, 2010. c  Springer-Verlag Berlin Heidelberg 2010