Dissimilarity representations in lung parenchyma classification

A good problem representation is important for a pattern re ognition system to be su essful. The traditional approa h to statisti al pattern re ognition is feature representation. More spe i ally, obje ts are represented by a number of features in a feature ve tor spa e, and lassiers are built in this representation. This is also the general trend in lung paren hyma lassi ation in omputed tomography (CT) images, where the features often are measures on feature histograms. Instead, we propose to build normal density based lassiers in dissimilarity representations for lung paren hyma lassi ation. This allows for the lassiers to work on dissimilarities be-tween obje ts, whi h might be a more natural way of representing lung paren hyma. In this ontext, dissimilarity is dened between CT regions of interest (ROI)s. ROIs are represented by their CT attenuation histogram and ROI dissimilarity is dened as a histogram dissimilarity measure between the attenuation histograms. In this setting, the full histograms are utilized a ording to the hosen histogram dissimilarity measure. We apply this idea to lassi ation of dierent emphysema patterns as well as normal, healthy tissue. Two dissimilarity representation approa hes as well as dierent histogram dissimilarity measures are onsidered. The approa hes are evaluated on a set of 168 CT ROIs using normal density based lassiers all showing good performan e. Compared to using histogram dissimilarity dire tly as distan e in a k nearest neighbor lassier, whi h a hieves a lassi ation a ura y of 92.9%, the best dissimilarity representation based lassier is signi antly better with a lassi ation a ura y of 97.0 % (p = 0.046)