INTRODUCTION

Head motions continue to be a major source of artifacts in structural and functional brain MRI.^1-4 Compatibility with existing scan protocols and ease of adoption are important considerations when selecting a motion tracking method. In this work we present initial feasibility results of MoCoPad, a novel pneumatic sensor-based head motion detection and tracking system for MRI. The system comprises of head pad with a matrix of air pressure sensors, which replaces the standard head pad in an MRI scan, and allows completely sequence and coil agnostic, fast tracking of head motions without the need for navigators, head markers or camera systems. Here in, we describe the construction, setup, modeling and initial motion detection and head pose estimation results in phantom studies in a 3T scanner.

METHODS

Sensor Pad and Data Acquisition. The sensor pad (Figure 1A) contains a head coil sized silicone rubber pad with a 4 x 4 grid of air chamber sensors. The chambers are connected to a RF shielded box with 16 analog pressure sensors (ABPDANV060PGAA5, Honeywell Inc, IN, USA) using 10 feet long 3.175 OD plastic pneumatic tubes. 16 channel data from the pressure sensors are digitized by an Arduino microprocessor at a baud rate of 115200 bit/s (10 bits/channel, 160 bits/reading, 1.38 ms/reading)⁵. Data are recorded continuously by a computer outside the scanner room via a shielded USB connection.
Motion Modeling A calibration model is needed to relate the head motion associated pressure changes in the sensor cells to the degree of head rotation. In this work, we trained a Gaussian Process Regression (GPR)⁶ based deep learning model using a head phantom on the bench, outside the scanner (Figure 2A). An optical motion tracker (Microntracker, Claronav, Toronto, Canada) was used measure head pose while the phantom was rotated at 100 random angles about the foot-head axis, with simultaneous recordings of pressure data. 80 data points where used to train the model with k = 5 k-fold cross-validation. The remaining 20 points were then used as test data to evaluate the model.
Scanner Experiments : We performed two experiments on a 3T Philips Elition scanner with the 2-channel body transmit/receive coil to test the feasibility of the technique.
Motion Detection and Frame Rejection: In the first experiment, we tested the ability of the system to detect head motion, and guide data rejection. We performed two 40 dynamic, single shot EPI scans (FOV 240 mm², 2.4 mm³ voxels, 36 slices, R = 2, TR/TE = 2046/30 ms, Tacq = 87 secs), where the phantom was manually moved to an in-plane rotated position for a few dynamics and returned back to the initial (or close to) position. Motion corrupted volumes were identified from time-stamped sensor data retrospectively and removed from the time series. Temporal signal to noise (TSNR) over the volumes were then compared between data with and without motion frame rejection.
Motion Estimation: In this experiment, we tested the feasibility of motion estimation based on sensor pad data. We acquired single slice data with two sequences: a 2D GRE with FOV 256 mm², 2 x 2 x 4 mm³ voxels, TR/TE = 40/2.1 ms, flip = 12 deg, Tacq = 5 secs and a 2D multishot EPI with FOV 228 x 228 mm², 1 x 1 x 4 mm3 voxels, EPI fac 19, TR/TE = 2000/30 ms, Tacq = 56 secs. Data were collected in three phantom positions, with the first reference position aligned with the 0 degree bench trained position. Sensor pad data were used to estimate motions retrospectively and compared to motions estimated by image registrations in matlab. To derive motions about the center of the image, we identified the pivot point in moved vs reference difference images and transformed the sensor pad rotation (which are modeled off the bottom of the phantom) to in-plane rotation and translations about the center of the image.

RESULTS

The on-the-bench motion model calibration showed high accuracies of 0.18 and 0.32 degrees of RMSE for the training and testing data (Figure 2B). Figure 3 shows results from the motion detection experiment. The sensor pressure data clearly depicted the change in head pose and head rotations estimated from the timestamped data allowed easy identification and rejection of the motion corrupted volumes. TSNR values computed for the whole phantom where significantly higher with frame rejection (12.86, 10.68) compared to without (8.16 and 7.38) for the two scans. Figure 4 shows results of the motion estimation. Motions estimated from sensor pad were close to the motions estimated from image registrations for both GRE and mshEPI, with maximal absolute translation and rotation errors being 3.74 mm and 1.57 degrees respectively.

DISCUSSION

In this work we have presented feasibility results of a novel head motion tracking system that is based on monitoring pressure changes caused by a moving head. The system does not interfere with the MRI sequence, does not impact patient comfort and is fast (1.38 ms/motion update) allowing for the potential prospective corrections by integrating with the scanner’s respiratory triggering system. Future work will focus on minimizing errors in motion estimation, developing fast in-scanner calibrations and developing prospective corrections.

Acknowledgements

This work was supported by NIH EB02558. We are very grateful to Dr William A Grissom and Dr Adam Anderson for valuable comments and advice.