Application of lung segmentation algorithm to disease quantification from CT images

Lung segmentation is a powerful tool in medical imaging. It can be used in quantification of the lung disease progression, regression or stagnation, and change in the visual extent of disease over time is an important marker of response to therapy and a predictor of mortality. In this work we present a lung disease quantification method that we applied on the baseline and follow-up lung CT images. We proposed a new method for disease progression based on standard segmentation and registration techniques. The main scope of the work is to allow the radiologists to measure volumetric changes of the lungs and to calculate the proportion of the functional reduction of the healthy parenchyma area of the patients with the pleural disease, mesothelioma, and other lung diseases. The proposed method compares the segmented area that is not affected by the disease inside the parenchyma of the original and the registered follow-up exam of the same patient. By calculating the area of the healthy tissue in parenchyma, we can conclude that: if the area of healthy tissue is larger than on the follow-up scan, that the disease progressed, otherwise, the disease regressed. Preprocessing of the images was done by registration and transformation of the images by affine and non-rigid B-spline method. Region growing algorithm is used for segmentation and by comparison of segmented structures from those images resulted in determining the percentage of disease progression, which we compared to visual observers. Total number of 15 patients was taken for testing; total of 1584 CT slices. For the verification of the results, Dice similarity coefficient is used for segmentation accuracy, and in order to compare the results of quantification of the disease with the manual findings, 3D volume of the lungs is measured for each patient and compared with the clinical findings. Pearson correlation coefficient shows significant correlation of our method with visual observers and that our method has high diagnostic accuracy.