Robust motion tracking in liver from 2D ultrasound images using supporters

Abstract

Purpose

Effectiveness of image-guided radiation therapy with precise dose delivery depends highly on accurate target localization, which may involve motion during treatment due to, e.g., breathing and drift. Therefore, it is important to track the motion and adjust the radiation delivery accordingly. Tracking generally requires reliable target appearance and image features, whereas in ultrasound imaging acoustic shadowing and other artifacts may degrade the visibility of a target, leading to substantial tracking errors. To minimize such errors, we propose a method based on so-called supporters, a computer vision tracking technique. This allows us to leverage information from surrounding motion for improving robustness of motion tracking on 2D ultrasound image sequences of the liver.

Methods

Image features, potentially useful for predicting the target positions, are individually tracked, and a supporter model capturing the coupling of motion between these features and the target is learned on-line. This model is then applied to predict the target position, when the target cannot be otherwise tracked reliably.

Results

The proposed method was evaluated using the Challenge on Liver Ultrasound Tracking (CLUST)-2015 dataset. Leave-one-out cross-validation was performed on the training set of 24 2D image sequences of each 1–5 min. The method was then applied on the test set (24 2D sequences), where the results were evaluated by the challenge organizers, yielding 1.04 mm mean and 2.26 mm 95%ile tracking error for all targets. We also devised a simulation framework to emulate acoustic shadowing artifacts from the ribs, which showed effective tracking despite the shadows.

Conclusions

Results support the feasibility and demonstrate the advantages of using supporters. The proposed method improves its baseline tracker, which uses optic flow and elliptic vessel models, and yields the state-of-the-art real-time tracking solution for the CLUST challenge.

Appendix: Baseline method: tracking by Makhinya and Goksel (TMG)

Our previously developed tracker [6], which is runner up of the Challenge in Liver Ultrasound Tracking (CLUST)-2015 challenge and is based on optic flow and elliptic vessel model, is employed as object tracker for tracking the supporters and target. The method is summarized below for completeness. Note that this method can track several landmarks together real time and works faster than US acquisition.

Overview The method decides in the initial frame, if the target is vessel-like or not by matching with ellipsoid vessel templates and integrates then several tracking strategies. It involves reference tracking (RT) when the local appearance on the initial, ${\mathbf {I}}^{0}$, and the current frame, ${\mathbf {I}}^{f}$, are similar. Meanwhile, it uses model-based iterative tracking (IT) when RT fails and local appearance of consecutive frames, ${\mathbf {I}}^{f-1}$ and ${\mathbf {I}}^{f}$, are similar. A robust motion tracking is applied in either case. For vessel-like structures, this is improved further by model-based tracking.

Motion tracking Lucas–Kanade-based tracking [20] was applied on a set of regularly spaced grid points around each target. RT is then used for exploiting the repetitive breathing motion characteristic, while IT is used for tracking the motion during the rest of the cycle, i.e., when RT fails. Each tracking strategy yields several motion vectors, which are then filtered for outliers. Finally, from the remaining motion vectors, an affine transform is computed to provide a robust motion estimate for the target.

Model-based tracking For vessel-like structures, model-based tracking is done using an axis-aligned ellipse representation of vessels. For each frame ${\mathbf {I}}^{f}$, first the center is transformed by the affine transform determined by motion tracking; see above, and then the center and radii are re-estimated as in [21] using the Star Edge detection, dynamic programming, model fitting, and binary template matching. The center of the resulting ellipse is then used as the estimated target position at frame ${\mathbf {I}}^{f}$.