Synopsis

Keywords: MR Fingerprinting/Synthetic MR, Quantitative Imaging, Gadgetron, Brian Tumor, MNI, FSL, Docker, Automated, Registration

In this study, to overcome some of the clinical integration issues of MRF, we present a fully automated online reconstruction and post-processing pipeline for 3D-MRF where the quantitative maps and custom reports are returned to the scanner in real time. The whole pipeline is hosted in a Kubernetes cluster which includes Gadgetron, FSL and other custom tools in discrete Docker images. To illustrate the capability of the pipeline, 3D-MRF raw datasets of healthy and brain tumor patient datasets are reconstructed in a cloud-based Gadgetron, registered to MNI space using FSL and analyzed to compare with population based regional maps.

Purpose

Although 3D Magnetic Resonance Fingerprinting (3D-MRF) allows for rapid acquisition of quantitative parameter maps, the process of preparing reconstructed images for radiologists to interpret is usually far slower, with multiple offline manual steps. As a result, fingerprinting maps are rarely available to radiologists via PACS - let alone while a subject is still in the scanner. Contributing to the issue, offline reconstructions often lack crucial header information needed for intrasession and intramodality image comparisons. These limits are particularly frustrating for MRF, which should natively support immediate intra- and intersubject comparisons that would be valuable both at the scanner and in retrospective studies - especially if maps registered to common atlases were also available.

To address these limits, we implement an online fully-automated pipeline capable of reconstructing 3D-MRF data and returning T1, T2, M0, and B1 maps in both subject and MNI-152 space while remaining fully DICOM-compatible. The immediate availability of quantitative maps in MNI-152 space is then leveraged to generate statistical reports comparing major brain regions against previously-collected population data via the Harvard-Oxford subcortical atlas while the subject is still in the bore.

Methods

In order to make a modular, scalable system, Docker images supporting ISMRMR-RD pipe-based communication were created and hosted within a remote Kubernetes cluster consisting of CUDA-capable GPU nodes. Cluster architecture and data flow details are available in Figure 1. All raw data and intermediate module outputs are uploaded to a subject-specific cloud blob container for future analysis and to ensure reproducibility.

Acquisition raw data is first sent to the Kubernetes cluster via the Framework for Image Reconstruction Environment (FIRE) prototype routed through an SSH jump node. Data is passed to the first module, a Gadgetron pipeline performing GPU-accelerated 3D-MRF reconstruction. The resulting T1, T2, M0, and B1 maps are then piped to the next module.

The second module is a Gadgetron pipeline generating synthetic MPRAGE contrasts via a previously-trained Tensorflow regression model. The pipeline returns synthetic T1w contrasts that closely match the MNI-152 T1w reference image. Both the original quantitative maps and synthetic MPRAGE contrast are then piped to the next module.

The third module is a python container running FSL via the nipype interface. First, incoming data is converted to NIFTI . Using the synthetic MPRAGE contrast, MRF maps are nonlinearly registered to MNI-152 using FLIRT and FNIRT. The resulting MNI-space maps and associated forward/inverse transformations are then piped to the next module.

The final module is a python container performing intra- and inter-subject regional analysis in MNI-152 space. Historical subject-specific or population data is retrieved from blob storage, and regional characteristics are compared against the current dataset to generate a summary report. The MRF quantitative maps in both subject-space and MNI-space as well as the regional statistics report are then piped back to the scanner console where they can be viewed, exported to DICOMs, or sent to a PACS database.

Existing raw datasets from 3 healthy volunteers and 3 brain tumor patients were retroreconstructed via FIRE to test the proposed system. All datasets were originally acquired with a prototype FISP 3D-MRF sequence on a 3T scanner (MAGNETOM Vida, Siemens Healthcare, Erlangen, Germany) with varying field of view and spatial resolutions. Postprocessing times for each module were recorded to measure the delay in image availability at the scanner. For the regional statistics reports, we chose to generate subject-vs-population regional comparisons as a proof of concept, with the population regions defined via aggregation of 60 MRF datasets from 10 healthy volunteers.

Results

Figure 2 demonstrates the quality of the MRF quantitative maps and synthetic MPRAGE images in both subject and MNI-152 space, as well as the quality alignment against the Harvard-Oxford subcortical atlas. Figure 3 are examples of the automated statistical report generated for the three healthy subjects. Figure 4 replicates Figure 3 for the three brain tumor patients, demonstrating the regional analysis’ sensitivity to atypical tissue properties across regions. Figure 5 summarizes the processing times of the various and the resulting delay in image availability at the scanner console

Discussion

A fully-automated online analysis pipeline is presented for 3D-MRF, capable of returning quantitative T1, T2, M0, and B1 maps in both subject-space and MNI-152 space as well as automatic intra- or inter- subject statistical reporting using MNI-space atlas regions.

Images and reports are returned to the scanner in real-time, and both raw and registered quantitative maps are fully compliant with common DICOM and NIFTI tools like FSL and 3DSlicer. Images and reports can also be saved to PACS directly, hopefully simplifying clinical adoption for radiologists and other clinicians.

Automated generation and semantic storage of MNI-space quantitative maps for each reconstructed dataset should substantially simplify future population-based research methods. For example, MNI-space images allow for immediate comparison against a subject’s past scans or even against other subjects can be performed without any additional registration, ROI creation, or other manual intervention.

Additionally, while still requiring substantial validation, the fully-automated generation of regional reports represents an initial attempt at simplified anomaly detection while still at the scanner console. We hope these reports can eventually serve as an “early warning” about potential anomalies so that technicians can acquire additional imaging immediately, avoiding a repeat visit.

Acknowledgements

This work was supported by Siemens Healthineers and Microsoft Corporation