Automated discrimination and quantification of idiopathic pulmonary fibrosis from normal lung parenchyma using generalized fractal dimensions in high-resolution computed tomography images

We computed generalized fractal dimensions for high-resolution computed tomography (HRCT) images to investigate their value in the discrimination and quantification of idiopathic pulmonary fibrosis (IPF) from normal lung parenchyma. A probability distribution that was based on the pixel value in each image was used to compute capacity, information, and higher fractal dimensions for a series of 52 HRCT slices obtained from four patients. Qualitative classification of normal, mild, moderate, and severe IPF cases was achieved by computing the following parameter: DD = D 0 − 2 D 1 + D 2, where D 0, D 1, and D 2 represents the capacity, information, and pair correlation dimensions, respectively. A multiple linear regression analysis using morphometric quantification for the set of 52 slices was tested for all possible combinations of the parameters D 0, D 1, D 2, and D 3. The generalizability of the model was tested by predicting the extent of IPF for each patient from a regression model computed with the remaining slices in the database. The best regression results were obtained using the independent parameters D 1 and D 2 to quantify the extent of diseased lung parenchyma. The technique was tested with 48 slices from 12 new patients. The results indicated that the extent of IPF could be predicted within the confidence limits given by the regression analysis. The extent of IPF can be predicted well within the 90% confidence interval given by the model. The width of the confidence interval decreases as the number of slices used in the linear regression model increases. This operator-independent quantitative technique may be useful in the follow-up of patients with IPF.