Two-Phase Supervised Segmentation Algorithm for Automatic Segmentation of Lung Parenchyma from Chest CT

摘要

Segmentation of lung parenchyma is a challenging task in the Computer Aided Diagnosis (CAD) of lung disorders using chest Computed Tomography (CT). In this research, a Two Phase Supervised algorithm has been proposed for segmentation of lungs in chest CT slices. In the first phase, the initial lung region is obtained by applying a combination of iterative thresholding and morphological operations. The shape features of the resulting lung region are applied to a decision tree classifier that is constructed from a training dataset to determine whether the segmented lung forms a complete lung. In the second phase, if the initial lung is complete the lung region is filled with lung tissue if the initial lung is not complete, the lung region is determined by a series of operations. First, the longest of the two connected components is determined. The longest connected component is then folded and translated horizontally. The two lung regions are then converted to a single connected component and the convex hull is obtained. The convex hull is interpolated to obtain the outer convex edge. The outer convex edge thus obtained is superimposed on the binary image obtained by folding and translation and used as the initial contour for the Active Contour Model (ACM). The ACM algorithm is iterated until the distance between the contours of two subsequent iterations becomes lesser than a threshold. It is also ensured that the number of components does not exceed two. This method is adaptive in that the number of iterations of ACM is not fixed and is based on the image for which it is applied. This method of lung segmentation has been compared with the conventional Iterative Thresholding Method, Convex Hull Based algorithm and Supervised algorithm for segmentation. The maximum overlap achieved with all the four methods is 100% while the minimum achieved with the proposed method is 55.3%, conventional iterative thresholding method is 37.83%, Convex Hull Based algorithm is 25.82% and Supervised algorithm is 54.25%. Thus, the proposed Two-Phase Supervised Method is found to be better than the other three methods with which the comparison is done.