1350568 - Clinical evaluation of two inline image reconstruction pipelines for compressed sensing-accelerated, first-pass cardiac perfusion MRI

Compressed sensing (CS) provides a means to achieve higher spatial resolution and/or extended myocardial coverage per unit time than conventional parallel imaging [1]. For clinical translation of said technology, one must consider clinical workflow (e.g., efficiency, automation, and integration). In this study, we sought to implement and compare two inline reconstruction pipelines for our 6-fold accelerated perfusion pulse sequence; specifically, we will evaluate the performance of “Yarra” and Siemens “Framework for Image Reconstruction Environments” (“FIRE”) frameworks.

Methods:

Pulse Sequence: We performed timing benchmarks at 1.5T (MAGNETOM Aera, Siemens Healthcare) using a 6-fold accelerated pulse sequence [1] designed to sample 4 slices per heartbeat. Relevant imaging parameters are FOV=384×384 mm², matrix size=192×192, 30 coils, 100 temporal frames, and raw data file size 1,264 MB. A phantom experiment was conducted 5 times for each framework.

Image reconstruction: CS was used to reconstruct the dataset. The code was GPU-accelerated in Matlab (NVIDIA Tesla V100, 16GB VRAM). The reconstruction method was implemented in both Yarra (yarra-framework.org) and FIRE platforms. For FIRE, we also included zero-filled (ZF) image reconstruction every two heartbeats with low latency for real-time display of gadolinium uptake.

Timing benchmarks: For Yarra, we included the following processing times: manual process to send the full k-space raw dataset to the Yarra server using a client (30sec), raw data transfer to the GPU server, and CS reconstruction. For FIRE, we included raw data transfer to the GPU server, which is streamed one k-space line at a time, and CS reconstruction (see Figure 1).

Clinical Translation: We performed clinical stress- and rest perfusion on 29 patients (16 men, mean age 67.34 y.) with a cardiac implantable electronic device (CIED) using our pulse sequence with inline image reconstruction.

Results:

Both FIRE and Yarra pipelines were successfully implemented and produced DICOM files automatically. Compared with Yarra (11min 49.8 ± 35.5sec), despite additional task to produce ZF images, FIRE (8min 26.8 ± 3.8sec) produced 28.6% faster CS reconstruction. Figure 2 shows stress and rest perfusion images of one representative patient with CIED demonstrating FIRE reconstruction; Yarra produced the same images (see Figure 3).  

Conclusion: This study shows that both FIRE and Yarra are useful platforms to perform inline reconstruction. Each platform has relative advantages and disadvantages. For smaller data sets such as 2D perfusion CMR, FIRE might be a better solution because it enables real-time display of gadolinium uptake and faster CS reconstruction. For larger datasets which take longer to reconstruct (e.g., non-Cartesian, 3D), Yarra might be a better solution because images can be sent back asynchronously while additional scanning is ongoing. Moreover, Yarra facilitates deploying algorithms to large fleets of scanners with different software versions.