Abstract: A challenge for computed tomography (CT) image analysis is to produce quantitative information of parenchymal abnormalities generalizable to a population of interest. Typically CT data from patient populations are heterogeneous, spatially correlated, massive and high dimensional. In order to classify abnormalities and their specific patterns across regions of interest (ROI) at the CT pixel level, it is important to normalize variation, utilize spatial information, and choose an efficient classification method. Decomposing CT images using a modified version of Aujol's algorithm, which was based on partial differential equations in negative Solobev space, facilitated normalization. Once decomposed, noise was removed to reduce variation, yielding a denoised image. Next, we used standard image texture features to capture spatial information within ROIs and to create the variables used to classify abnormalities. Abnormalities were classified via two methods, namely, multinomial logistic regressions (using backward selection and predicted probabilities) and non-concave penalized likelihood equations combined with support vector machine (SVM) classification. Initially we applied our techniques to classify patterns of Scleroderma-related lung abnormalities (ground glass, fibrosis and honeycombing) in small ROIs. Finally we extended the classification scheme to the whole lung, accounting for structure, in order to produce quantitative scores of sclerodermal abnormalities in lung parenchyma of patients.