Evaluating motion correction for structural brain scans

Siemens provided the MindCORE Neuroimaging Facility witha demo of their built-in motion correction software for structural scans (BioMatrix Motion Sensor – SAMER). Below are examples of T1 MPRAGE scans collected from a participant performing various commonly observed movements while being scanned. Whole brain MPRAGEs were collected with the following parameters: matrix 256 x 256; 192 sagittal slices; 0.9mm slice thickness; FOV 240 x 240mm; resolution .9 x .938 x .938; TR 2300ms; TE 2.32ms; flip angle 8°; acceleration factor 2. The scans were collected on a 3T Siemens Cima.X scanner (software version XA61) with a 32-channel head coil and with normal padding placed around the participant’s head to reduce extraneous head movements. All images are presented in native space and with no additional processing. The default sequence (sequence name: tfl3d1_228ns) produces an uncorrected and a motion-corrected version of each scan, allowing the user to evaluate the effectiveness of the motion correction algorithm.

Each of the scans below is presented with key raw scan slices as well as with motion correction enabled.

Baseline Scan

No intentional participant motion. 

Heavy Breathing

Participant breathing slowly and heavily with repetitive chest and some chin motion. 

Chewing

Participant moved jaw and swallowed many times during the scan.

Movement

Participant moved substantially during the scan by shifting their weight, and purposely pressing their head against the sides of the head coil and compressing the padding.

Freesurfer segmentation

Each of the above scans were processed using Freesurfer’s recon-all command with default parameters. Presented below are representative coronal slices showing the gray-matter/white-matter boundary (blue), the pial surface (red), and major subcortical structures shaded. Segmentation errors (such as can be seen in the right hippocampus) were not manually corrected. 

As can be clearly seen, segmentation largely fails in cases where there are major motion events (i.e., in the “move” condition) and these failures are not overcome with the motion correction procedure. Interestingly, segmentation appears worse with motion correction in the “breathe” condition. The motion correction procedure did not produce any differences in the “baseline” and “chew” conditions between the corrected and uncorrected output (see values in table below), meaning motion artifacts were not detected in these conditions. 

The following table presents volumes (in mm³) of segmented structures for each scan with and without motion correction. 

Recommendation

The Siemens BioMatrix Motion Sensor procedure appears to produce qualitatively better output when there are moderate amounts of movement, as in the “breathe” condition above. These apparent improvements may not be beneficial in automated segmentation pipelines, however. In cases where there is little motion (“baseline” and “chew” conditions), the motion correction algorithm does not appear to be engaged. The motion correction algorithm is not able to overcome artifacts in cases where there are large motion events (“move” condition). If such large motion artifacts are apparent on reconstructed scans, it is recommended to repeat the scan, perhaps with additional coaching or other support for the participant.

Acknowledgements: Thank you to Alex Gordienko for developing and implementing the behavioral protocol here, as well as for the reconstruction and display of the scan quality.