Supervised Texture Classification for Segmentation of Abnormal Lung in Ct

Delineation of lung fields in presence of diffuse lung diseases, such as interstitial pneumonias (IP), challenges conventional gray level based segmentation algorithms. To deal with IP patterns affecting lung borders, a texture classification scheme for lung segmentation is proposed. The proposed method is based on supervised texture classification to distinguish surrounding tissue (ST) from lung parenchyma (LP), both normal and IP affected, since IP patterns are primarily manifested as texture alterations. A support vector machine classifier was trained to distinguish the ST from LP classes based on second order statistics textural features extracted from a 9x9 sliding region of interest across a CT image. A case sample of 120 HRCT images depicting abnormal lung boundary, corresponding to 17 patients diagnosed with IP secondary to connective tissue diseases, was analyzed. The segmentation accuracy of the method was assessed by comparing automatically derived lung borders to manually traced ones, by an experienced radiologist, using area-overlap and border shape differentiation metrics. Average segmentation accuracy in terms of area-overlap was 0.940±0.028, while border shape differentiation with respect to mean, root mean square and maximum distance were 1.161±0.401, 1.520±0.832 and 5.200±3.752 mm, respectively. The method is envisioned as an initial step of a computer aided quantification scheme for IPs.