Purpose

Motion artifacts in MRI are usually accompanied by visible image artifacts. In quantitative MRI, results may be affected in a more subtle way: values in the parametric maps can be corrupted without obvious hints in the appearance of the maps¹.
Magnetic Resonance Fingerprinting (MRF)² has a certain inherent robustness with regard to motion due to the applied pattern matching approach2. While MRF is indeed fairly insensitive to in-plane motion, several works suggest that 2D MRF is somewhat sensitive to strong through-plane motion^3.4.
In this work, a method is presented to detect whether through-plane motion has occurred during a 2D slice-selective FISP-MRF⁵ experiment, purely based on analyzing the MRF signal course, i.e. without the need for external motion detection such as navigator scans or a camera.

Methods

We used a prototype implementation of FISP-MRF⁵ with a prescan-based⁶ B1+ correction and spiral sampling (undersampling factor 48, FoV 256 mm, resolution 1 mm, slice thickness 5mm) with a spiral angle increment of 82.5°⁷. 12 volunteers’ and 32 patients’ brains (suspicion of glioma) were scanned on 3T MRI scanners (MAGNETOM Prisma, MAGNETOM Prisma^fit and MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany).
The motion detection method assumes that:

• the bulk movement of the head during the scan is rigid
• signal alterations due to motion have high frequency compared to the dictionary signals
• only signal alterations in solid matter are considered since e.g. fluid motion cannot be considered rigid

In addition to the ordinary MRF reconstruction, the acquired signals were normalized and compressed by adding each subsequent 48 (=undersampling factor) time points together. The normalized fingerprints obtained by matching the signals to the dictionary were utilized as a ground truth reference and compressed in the same way. By subtracting the compressed acquired signals and compressed reference fingerprints, residual maps for each compressed time point were generated. The process is schematically depicted in Figure 1.
The residual maps from volunteer and patient scans were labelled manually as motion corrupted (type: nodding, tilting, or stretching; severity level 1,2, or 3) or not motion corrupted. An overall estimation of the motion impact that occurred in a slice can be calculated (non, low, medium, strong) from that.
In each of the 32 patients, at least 10 slices were scanned, once with FISP-MRF using 3000 TRs and twice with 1500 TRs. The difference of parameter values in ROIs in solid matter were calculated between the two 1500 TR acquisitions of the same slices and related to the detected motion. Across all patients and for each patient individually, the fraction of detected motion in the scanned slices was calculated.
Finally, a convolutional neural network was implemented to replace the manual classification of spatiotemporal residuals automatically. It was trained by using the manual classification of volunteer data and the data of 24 patients and tested on the data of the other 8 patients.

Results

Three distinct patterns can be observed when analyzing the spatial residuals. Deviations dominated by gradients in anterior-posterior direction are caused by a nodding-like motion. Tilting-like movements lead to similar gradients in left-right direction, and stretching causes an offset over the whole brain. These patterns occur with different strengths and are exemplarily depicted in Figure 2.
Overall, no motion was detected in 77.0% of cases (i.e. slices), low motion in 8.8%, medium motion in 9.4%, and strong motion in 4.8% (Figure 3a)). This differs between unaffected scan sessions and strongly motion-corrupted ones (Figure 3b)).
Relative T1 and T2 differences in solid-matter ROIs between repeated MRF scans using 1500 TR are depicted in Figure 4. The differences are grouped for cases with no motion in one repetition and in the other repetition also none, low, medium, or strong motion. The relative differences scale with the strength of the detected motion.
The neural net had an accuracy of 94.2% on single residual maps. The overall estimated motion per slice by the neural network and the corresponding manual assessment for two patients are presented in Figure 5.

Discussion & Conclusion

By analyzing spatiotemporal residuals, the impact of motion can be identified which is not reflected by similarity measures over the whole signal course such as the inner product. In consequence, this allows to highlight potential inaccuracies in the quantitative data. Our proposed approach is facilitated by the fact that brain motion can be considered as predominantly rigid.
This novel method was successfully validated by comparing relative changes of T1 and T2 values to detected motions in repeated measurements. Results in a clinical study show that about 15% of the scanned data in patients were motion corrupted, while in some patients almost all data were corrupted and in some only a few slices.
Bulk motion artifacts can be superimposed by e.g. flow artifacts which makes simple fitting approaches unsuitable and requires to consider large brain area rather than single pixels. An approach to detect complex motion patterns is deep learning. The employed neural net had a high accuracy and was able to reproduce human estimations.
Future work is envisioned to provide more thorough validation by involving recorded motion data via e.g. an in-bore camera or artificially shifting and angulating the slice during the scan while the object/subject is kept still.

Acknowledgements

No acknowledgement found.