3D lung tumor motion model extraction from 2D projection images of mega-voltage cone beam CT via optimal graph search

In this paper, we propose a novel method to convert segmentation of objects with quasi-periodic motion in 2D rotational cone beam projection images into an optimal 3D multiple interrelated surface detection problem, which can be solved by a graph search framework. The method is tested on lung tumor segmentation in projection images of mega-voltage cone beam CT (MVCBCT). A 4D directed graph is constructed based on an initialized tumor mesh model, where the cost value for this graph is computed from the point location of a silhouette outline of projected tumor mesh in 2D projection images. The method was first evaluated on four different sized phantom inserts (all above 1.9 cm in diameter) with a predefined motion of 3.0 cm to mimic the imaging of lung tumors. A dice coefficient of 0.87 +/- 0.03 and a centroid error of 1.94 +/- 1.31 mm were obtained. Results based on 12 MVCBCT scans from 3 patients obtained 0.91 +/- 0.03 for dice coefficient and 1.83 +/- 1.31 mm for centroid error, compared with a difference between two sets of independent manual contours of 0.89 +/- 0.03 and 1.61 +/- 1.19 mm, respectively.