Tracking Lung Tumors in Orthogonal X-Rays

Abstract

This paper presents a computationally very efficient, robust, automatic tracking method that does not require any implanted fiducials for low-contrast tumors. First, it generates a set of motion hypotheses and computes corresponding feature vectors in local windows within orthogonal-axis X-ray images. Then, it fits a regression model that maps features to 3D tumor motions by minimizing geodesic distances on motion manifold. These hypotheses can be jointly generated in 3D to learn a single 3D regression model or in 2D through back projection to learn two 2D models separately. Tumor is tracked by applying regression to the consecutive image pairs while selecting optimal window size at every time. Evaluations are performed on orthogonal X-ray videos of 10 patients. Comparative experimental results demonstrate superior accuracy (~1 pixel average error) and robustness to varying imaging artifacts and noise at the same time.

A constrained linear regression optimization algorithm for diaphragm motion tracking with cone beam CT projections

“…For rotational projection images, depending on the view angles, complex anatomies such as vertebrate body and tissue could be overlaid on top of the tumor target, making the reliable target identification practically impossible. Even with sophisticated feature recognition techniques it is still hard to extract the tumor motion from the projection images or orthogonal X-rays, despite the relatively long computing time [28,29]. The detection of tumor motion is easier on AP image sequences than on rotational projection images.…”