Dynamic online registration guided collaborative tracking of lung tumor movement in fluoroscopic videos

Abstract

Tracking of the lung tumor movement in fluoroscopic video sequences is clinically significant and challenging problem due to the blurred appearance, sternum occlusion, and complicate intra- and inter- fractional motion. This introduces landmark ambiguity for accurate contour tracking. As the boundary of the lung, or part of the lung, is usually clear and can be accurately tracked, we propose a novel method to compute an accurate prior for the tumor based on the motion registration of the lung. For adapting to the appearance changes due to motion and reducing the label of annotation, we propose to apply online updated collaborative trackers to refine the boundary of the tumor. This motion registration guided online collaborative tacking algorithm is proven to be successful in real clinical dataset, especially for cases with unclear tumor boundaries. Excellent results are obtained on twelve motion sequences which contains 3531 frames in total.