Towards Markerless Optical Tracking for Prospective Motion Correction in Brain Imaging
Julian Maclaren1, Andre Kyme2,3, Murat Aksoy1, and Roland Bammer1

1Department of Radiology, Stanford University, Stanford, CA, United States, 2Department of Biomedical Engineering, University of California Davis, Davis, CA, United States, 3Brain and Mind Centre, University of Sydney, Sydney, Australia

Synopsis

Prospective motion correction based on optical tracking shows promise for improving image quality in MR brain imaging. To simplify this technique and expedite clinical deployment, it is desirable to avoid attaching markers to the patient’s head. Here we demonstrate proof-of-principle markerless tracking using an MR-compatible stereo camera and head coil configuration. We tested the method outside the MR environment using a 6-axis robot, capable of very accurate and repeatable (~20 µm) motion, to control a head phantom. Close agreement between our pose estimates and the applied motion suggests that accurate markerless tracking of the head is feasible in MRI.

Introduction and Purpose

Prospective motion correction has been shown to be effective in preventing motion artifacts in neuroimaging (1). Optical systems are a popular means of obtaining the required head tracking information during scanning (2–6). However, one challenge shared by all existing optical prospective motion correction methods is the need to attach markers to the subject. The aim of this work was to investigate the feasibility of markerless optical motion correction for neuroimaging using a realistic test platform with controlled motion.

Methods

Mechanical setup: The test platform (Figure 1) was based around a 6-axis robot arm (C3-A601ST, Epson America Inc.) capable of highly repeatable (20 μm) arbitrary motion in six degrees of freedom. An 8-channel head coil (Invivo) was securely mounted in front of the robot. A polycarbonate rig was designed and 3D printed to fit the head coil and rigidly hold two MR-compatible cameras normally used at our institution for marker-based optical head tracking. We used a mannequin head attached to the robot end effector to provide a realistically shaped surface for tracking. To achieve a skin-like surface for tracking, an image from Subject #5 of the MIT-CBCL face recognition database (7) was color-printed on paper and glued to the forehead (Figure 1b).

Calibration: Stereo camera calibration was performed by programming the robot to move a checkerboard marker composed of an 8 x 7 grid of 4 mm squares (Figure 2) to 30 different positions within the field-of-view of both cameras. Intrinsic and extrinsic camera parameters were then computed using the Matlab Calibration Toolbox (8).

Motion experiment: The robot was programmed to move between a series of fixed poses comprising translations in the head-feet direction and rotations about the z-axis (‘head shaking’). At each pose, synchronized images from the cameras were saved for offline motion processing.

Analysis: The feature-based pose estimation method reported previously for PET imaging of rats (9) was adapted for this experiment. In this method, features are detected and matched across multiple camera views to accumulate a database of head landmarks; pose is then estimated based on 3D-2D registration of the landmarks to features in each image. Estimates were compared to the ground truth motion applied by the robot.

Results

Figure 3 shows camera images of the mannequin head surface before (Fig. 3a) and after (Fig. 3b) masking and detection of SIFT (scale invariant feature transform) features. Figures 4 and 5 compare measured translations and rotations with the ground truth motion provided by the robot. The mean absolute error in pose detection was 0.14 mm and 0.23 degrees, for translations and rotations respectively.

Discussion

The aim of this work was to demonstrate the feasibility of accurate motion tracking of the head without attached markers, by using a test setup that is both realistic and allows for motion to be accurately specified. Initial results indicate that the geometric configuration and pose estimation algorithms can provide pose estimates suitable for sub-millimeter / sub-degree motion correction.

Since the head coil and camera setup used in this work are similar to the marker-based setup we currently use for in vivo marker-based prospective motion correction, translating the markerless method to in vivo applications would be straightforward. Other markerless tracking approaches based on structured light have been demonstrated previously for PET imaging (10). Structured light methods may require a larger field of view for accurate topological mapping of the head and are therefore less likely to be compatible with enclosed head coils used in MRI. The approach presented here does not suffer from this problem. However, poses are currently computed offline and in order for the method to be applied in prospective motion correction improved computational efficiency must be improved to allow real time pose processing with low latency.

Conclusion

We have tested a markerless head motion tracking technique using a realistic test platform that allows highly reproducible motion. Accurate motion tracking of a skin-like surface without the need for attached markers has been demonstrated.

Acknowledgements

NIH (2R01 EB002711 , 5R01 EB008706, 5R01 EB011654), the Center of Advanced MR Technology at Stanford (P41 RR009784), Lucas Foundation. Kyme is supported by funding from the Education and Research Foundation, USA Society of Nuclear Medicine and Molecular Imaging.

References

1. Maclaren J, Herbst M, Speck O, Zaitsev M. Prospective motion correction in brain imaging: A review. Magn Reson Med 2013;69:621–636. doi: 10.1002/mrm.24314.

2. Zaitsev M, Dold C, Sakas G, Hennig J, Speck O. Magnetic resonance imaging of freely moving objects: prospective real-time motion correction using an external optical motion tracking system. Neuroimage 2006;31:1038–1050.

3. Qin L, van Gelderen P, Derbyshire JA, Jin F, Lee J, de Zwart JA, Tao Y, Duyn JH. Prospective head-movement correction for high-resolution MRI using an in-bore optical tracking system. Magn Reson Med 2009;62:924–934. doi: 10.1002/mrm.22076 [doi].

4. Aksoy M, Forman C, Straka M, Skare S, Holdsworth S, Hornegger J, Bammer R. Real-time optical motion correction for diffusion tensor imaging. Magn Reson Med 2011;66:366–378. doi: 10.1002/mrm.22787.

5. Schulz J, Siegert T, Reimer E, Labadie C, Maclaren J, Herbst M, Zaitsev M, Turner R. An embedded optical tracking system for motion-corrected magnetic resonance imaging at 7T. Magn. Reson. Mater. Physics, Biol. Med. 2012:1–11. doi: 10.1007/s10334-012-0320-0.

6. Maclaren J, Armstrong BSR, Barrows RT, et al. Measurement and Correction of Microscopic Head Motion during Magnetic Resonance Imaging of the Brain. PLoS One 2012;7. doi: ARTN e48088DOI 10.1371/journal.pone.0048088.

7. Weyrauch B, Heisele B, Huang J, Blanz V. Component-Based Face Recognition with 3D Morphable Models. 2004 Conf. Comput. Vis. Pattern Recognit. Work. 2004:0–4. doi: 10.1109/CVPR.2004.41.

8. Bouguet J-Y. Complete Camera Calibration Toolbox for Matlab. Jean-Yves Bouguet’s Homepage 1999. http://www.vision.caltech.edu/bouguetj/calib_doc/

9. Kyme A, Se S, Meikle S, Angelis G, Ryder W, Popovic K, Yatigammana D, Fulton R. Markerless motion tracking of awake animals in positron emission tomography. IEEE Trans. Med. Imaging 2014;33:2180–90. doi: 10.1109/TMI.2014.2332821.

10. Olesen OV, Paulsen RR, Højgaard L, Roed B, Larsen R. Motion tracking for medical imaging: A nonvisible structured light tracking approach. IEEE Trans. Med. Imaging 2012;31:79–87. doi: 10.1109/TMI.2011.2165157.

Figures

Figure 1: Test platform: (a) Epson C3 6-axis robot to control the pose of a mannequin head inside an 8-channel head coil; custom-designed rig securing two MR-compatible cameras for prospective motion correction. (b) Styrofoam head phantom with ‘skin’ patch, allowing the setup to be tested using realistic geometry and texture.

Figure 2: (a) Checkerboard marker attached to the robot end effector for the one-time extrinsic camera calibration. (b) View of calibration pose #1 from both cameras. A total of 30 poses were used for the extrinsic calibration.

Figure 3: (a) Synchronized images of the head phantom from Camera 1 (left) and Camera 2 (right); (b) the same images showing matching SIFT features, after application of a mask to remove any influence of the stationary head coil.

Figure 4: Ground truth translation (blue crosses) as performed by the robot, plotted against motion estimated using the markerless system (red crosses). The mean absolute error in the estimated poses relative to the ground truth was 0.14 mm.

Figure 5: Ground truth rotation (blue crosses) performed by the robot, plotted against motion estimated using the markerless system (red crosses). The mean absolute error in the estimated poses, relative to the ground truth was 0.23 degrees.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1858