Quantifying Repeatability of a 3D-MRF Protocol During Scanner Software Upgrades
Andrew Dupuis1, Yong Chen2, Mark A Griswold1,2, and Rasim Boyacioglu2
1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 2Radiology, School of Medicine, Case Western Reserve University, Cleveland, OH, United States

Synopsis

Motivation: Magnetic Resonance Fingerprinting relies on the stable performance of vendor hardware and software through-time. However, the effects of software upgrades on the repeatability of quantitative protocols require evaluation.

Goal(s): Measure changes to MRF-derived T1 and T2 values resulting from a vendor-provided software upgrade.

Approach: 8 healthy volunteers were imaged before and after a software upgrade. Regional T1 and T2 changes were evaluated using correlation plots and MNI-derived regional means.

Results: The evaluated sequence was similarly repeatable in both the intra-version test-retest case (T1: -0.49+/- 3.85%, T2: -0.53+/-7.59%) and pre- to post-upgrade (T1: -0.01+/-4.23% , T2: -2.10+/-8.44%).

Impact: Showing similar repeatability of 3D MRF maps before and after vendor software upgrade establishes performance reliability. The automated analysis pipeline does not require qualitative images and can be used to test MRF reproducibility in various stages of longitudinal studies.

Introduction

Clinical or large scale research deployment of Magnetic Resonance Fingerprinting (MRF) based techniques relies on the stability of the measurement system and protocol, regardless of changes to scanner hardware and software [1,2]. However, logs of changes made by a vendor to scanner software defaults, research APIs, or hardware configurations are largely unavailable, which may result in affected relaxation maps [3]. We perform a precision comparison of a clinically deployed 3D-MRF sequence [4] before and after upgrading to a new scanner software baseline.

Methods

11 healthy volunteers were imaged after obtaining consent according to our institutional IRB using a 3T scanner hardware before a software upgrade (VA30 to VA50, MAGNETOM Vida, Siemens). Data from 8 volunteers, which returned for the second scan and did not have motion artifacts, were retained. All data was acquired using mrftools[cite] and stored in the ISMRM-RD format, as well as reconstructed online using the FIRE prototype interface [5] with additional acquisition timing details (TR, TE, Flip Angle, RF Phase, and Readout ID) logged to user-defined fields within the acquisition header. A hash of these acquisition headers was used to verify that sequence timing and acquisition settings remained consistent at runtime, after the software upgrade. Raw data structure was also characterized to ensure data arrays were populated in the same manner.

Volunteers were scanned twice during each session, on each scanner version to establish the test-retest versus pre/post-upgrade repeatability of the imaging method. The repeatability of both overall correlation and brain region mean was tested for both before and after the software update to ensure the repeatability was truly scanner software agnostic. All MRF datasets were registered linearly to the pre-upgrade initial acquisition using a synthetic MPRAGE lookup table [cite abstract?] and the FLIRT image registration toolbox [6,7]. After the registration, a single nonlinear registration was performed to generate voxel masks for brain regions based on the MNI-152 Harvard-Oxford Subcortical Structural Atlas. Figure 1 depicts the pipeline used for image registration and evaluation.

Results

When we ran the sequence for the new version that was cross-compatible for multiple previous versions, image quality was noticeably degraded. After working with the vendor to pinpoint the change made in the software upgrade, the issue was resolved but the study had to pause for two months. 32 imaging sets of data from 8 volunteers were converted to ISMRM-RD format for future analysis. Figure 2 demonstrates the correlation consistency of aggregate T1 and T2 values across the target population between pre- and post-upgrade data. Figures 3 and 4 compare the reproducibility of 3D-MRF acquisitions using regional means compared between immediate repetitions and multiple repetitions on two scanners respectively. Figure 5 summarizes the mean percent differences and standard deviations of 3D-MRF before and after the upgrade: in the intra-version test-retest case within the same scan session, -0.49+/- 3.85% variability was observed in T1, with -0.53+/-7.59% variability in T2. In the cross-version case, -0.01+/-4.23% variability was observed in T1, with -2.10+/-8.44% variability in T2.

Discussion

The repeatability of the examined 3D-MRF sequence was found to be very similar whether an examination was repeated immediately or repeated more than 2 months after a software version upgrade across all subjects. Good correlation was seen in both T1 and T2 values from all examined brain regions. Importantly, the proposed pipeline does not require qualitative images for registration between imaging sets or non-linear registration to MNI space which typically is based on T1-weighted images such as MPRAGE.. The pipeline can be deployed at various stages of longitudinal studies to confirm baseline repeatability and reproducibility after hardware or software updates. Confirmation of reproducibility after software or hardware upgrades is inherently important for confidence in data acquisition, an attribute that may have been taken for granted or assumed in previous updates.

While identical sequence source code on previous software versions from our vendor (VA20/VA30/VA31; Siemens) on multiple hardware platforms (Vida/Skyra/Aera; Siemens) had yielded identical headers and raw data formatting, initial compilation of the same codebase for the VA50 platform did not yield the same hash or raw data behaviors. This difference in headers is due to API changes within the vendor’s acquisition, raw data management, and image reconstruction systems. With vendor assistance, modifications to the sequence were made until sequence hash and raw data structure matched the VA31 system, emphasizing the importance of cooperation between researchers and vendors to ensure consistent performance across software upgrades.

Acknowledgements

No acknowledgement found.

References

[1] Lo et al. Multicenter Repeatability and Reproducibility of MR Fingerprinting in Phantoms and in Prostatic Tissue. Magn Reson Med. 2022 Oct;88(4):1818-1827. doi: 10.1002/mrm.29264. Epub 2022 Jun 17. PMID: 35713379; PMCID: PMC9469467.
[2] Körzdörfer et al. Reproducibility and Repeatability of MR Fingerprinting Relaxometry in the Human Brain Radiology 2019 292:2, 429-437
[3] Keenan et al. Assessing changes in MRI measurands incurred in a scanner upgrade: Is my study comprised?
[4] Dupuis et al. Quantifying 3D-MRF Reproducibility Across Subjects, Sessions, and Scanners Automatically Using MNI Atlases. ISMRM 2023.
[5] Chow K, Kellman P, Xue H. Prototyping Image Reconstruction and Analysis with FIRE. SCMR 2021.
[6] Jenkinson et al. Improved Optimisation for the Robust and Accurate Linear Registration and Motion Correction of Brain Images. NeuroImage, 17(2), 825-841, 2002
[7] Woolrich et al. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45:S173-86, 2009

Figures

Figure 1: Repeatability analysis pipeline. Three sets of transformations bring atlas regions to individual 3D-MRF spaces. Initial Synthetic lookup table application results in the MPRAGE-like contrast that the FSL relies on. Linear transformation between sets is done as the second step with FSL FLIRT. Finally, original synthetic MPRAGE goes through non-linear registration to MNI space. Atlas label maps are generated in the individual set spaces via the inverse of the transformation matrices.

Figure 2: Correlation of mean regional T1 and T2 values between results pre- and post-upgrade. All the regions are distributed along the identity line.

Figure 3: Repeatability with test-retest intra-version. Bland-Altman plots show the baseline variation in repeat scans over the two versions. The distributions are similar over regions and volunteers.

Figure 4: Reproducibility with test-retest inter-version. Mean and standard deviations are similar to Figure 3 indicating that the software upgrade did not cause additional variability.

Figure 5: Aggregate T1 and T2 repeatability and reproducibility over all the volunteers and regions.