Prospective motion correction for MRI neuroimaging has become increasingly prevalent in practice, with a variety of techniques and hardware being actively developed¹. Reliable motion estimation is crucial for prospective correction, since incorrect estimation of the motion will degrade the image.

Tracking can be accomplished using MR navigators or external devices. MR navigators often prolong the scan time, lack simultaneous 3D spatial and temporal resolution, and are not compatible with all image acquisitions. External tracking systems using markers have difficulties validating marker attachment and as a result the estimated brain motion is uncertain. Furthermore, these methods require extra patient preparation for attaching the markers, which also makes them inconvenient for clinical use.

In the search for a clinically feasible method, we have developed a markerless tracking system, and previously demonstrated it for retrospective correction². In this work we demonstrate prospective motion-correction in a 3D FLASH sequence using our system. Markerless tracking is non-trivial and can be computationally demanding, but here we demonstrate that it can be successfully implemented for real-time prospective correction in MRI.

Our tracking system was set up on a hybrid 3T mMR Biograph, Siemens³. The control unit of our system was connected to the internal scanner Ethernet network. A healthy volunteer was scanned using a custom 3D FLASH sequence, modified to query our control unit for the current patient position⁴. Changes in orientation and shifts in position were corrected for in real time by dynamically adjusting the gradients, RF pulse frequencies, and k-space sample phase every TR, thus generating a corrected image on the scanner upon completion of the scan.

One volunteer, having given informed consent, was instructed to perform a repeatable motion pattern or remain motionless while being scanned with and without motion correction enabled. The parameters of the 3D FLASH sequence were: matrix 128x128, pixel spacing 1.72 mm, 56 slices, slices thickness 3 mm, TR 30 ms, TE 2.2 ms, and flip angle 35 degrees. A queue lag of pose estimates of 4x TR was maintained by the scanner to prevent correction queue underflow. This lag prolonged the latency between performed motion and motion update in addition to the pose estimation and communication latency.

A standard MPRAGE image volume (512x512x192, pixel spacing 0.49 mm, and slice thickness 1 mm) was used to determine the geometric alignment between the tracking system and the scanner coordinate system. A rigid transformation was fit iteratively between a surface of the volunteer’s face obtained from our tracking system, and the surface of the volunteer’s head extracted from the MPRAGE volume. The alignment transformation was subsequently used to transform the volunteer’s position into the scanner coordinate system. No further calibration of the system was required.

Figure 1 shows 3D point clouds of the volunteer inside the MR head coil. These constitute some of the outer positions, corresponding to the maximal motion shown in Fig. 2. Figure 2 shows the motion recorded during the image acquisitions with/without motion correction (moco) and the reference scan with no motion. The curves represent the motion of the centroid of the 3D point cloud. It is noted that the motion pattern is similar between the scans without and with motion correction. All of the plots have ripples corresponding to respiratory motion with amplitude of 0.2 to 0.5 mm (Fig. 3). This indicates a tracking accuracy substantially better than the general image resolution. Figure 4 shows an image slice from each of the three FLASH volumes. A much better result is obtained with motion correction than without. The slice with motion correction is comparable in quality to the slice with no motion.

The only correction applied was adaptively updating the imaging coordinates according to the tracked motion. Further steps such as B0 and B1 correction were not included, and these can be expected to improve image quality further, especially for large motions. The tested motions are relatively large demonstrating that this system is capable of correcting larger-than-usual ranges of motion.

Determining a position quickly and accurately are key challenges in tracking for prospective motion correction. Generally it is possible to trade speed for accuracy and vice versa. The presented results show that it is possible to find a good tradeoff between these two objectives with further optimization still possible.

We have shown that prospective motion correction can be implemented using markerless tracking. Markerless tracking has great potential for clinical routine as failures from maker attachment are eliminated, it does not introduce any extra patient discomfort or preparation time, and it also does not prolong the scan time.