Animating Terabytes of Imaging Data from a One-Minute Scan: Interactive Reconstruction of Flexibly Acquired MRI Data
David S Smith1, Saikat Sengupta1, Aliya Gifford1, and E Brian Welch1

1Institute of Imaging Science, Vanderbilt University, Nashville, TN, United States

Synopsis

We present a software system called CITRON that can extract, reconstruct, and display images from non-Cartesian MR data at 60 frames per second, allowing the user to vary the reconstruction parameters in real time in order to explore and optimize the reconstructed images interactively.

Purpose

The current paradigm of MR data collection and image archival is sub-optimal. Despite the high cost of MR exams, the data collected is fixed in an ad hoc image representation for storage in the database and the raw data discarded, along with any information it may still contain. New breakthroughs in MR reconstruction and acquisition coupled with faster hardware means that now MR data can be thought of as “Big Data” that can be explored on demand and used to answer a larger spectrum of clinical and research questions with greater breadth and accuracy.

Here we present CITRON1 (Command-driven Interactive TRajectory Optimized Nufft), a system to extract and reconstruct images from whole-body continuous golden angle radial MRI data at up to 60 frames per second (fps), allowing the user to interactively optimize the reconstructed images for greater clinical and research utility.

Methods

An adult male volunteer was scanned at 3 T under an IRB-approved protocol using a whole-body continuously moving table golden angle radial scan2 that covered 1.5 m in table distance at 2 cm/s. The raw data contained 8 channels of 512 readout samples by 21,207 radial profiles. The transverse in-plane field of view was 40 cm x 40 cm with a voxel size of 1.56 mm x 1.56 mm. TR/TE was 3.7/1.35 ms, flip angle was 20°, and total scan time was 75 s.

Figure 1 shows the two subsystems of CITRON: a user interface and a reconstructor. The user interface, driven by OpenGL, displays the reconstructed images and accepts user commands, which it passes to the reconstructor. Reconstruction is performed with TRON, a fast non-Cartesian reconstruction written in CUDA3 that receives commands and writes back a float RGBA image into an OpenGL texture. This happens within milliseconds so that parameters can be continuously varied while the image updates at movie-like frame rates of 20-30 fps or more, depending on hardware.

Results

An example CITRON window is shown in Figure 2. The image is shown over the whole window with an informational text overlay. The command set is at upper right. The current parameters are across the bottom. Currently, the user can vary the image size, slice position down to the individual profile (70 micron resolution in this data), slice thickness down to the profile, gridding kernel width and oversampling factor, and coil combination method (sum-of-squares or adaptive4). Adjustability depends on the acquisition: golden angle radial is more flexible than Cartesian, for example.

In the window title in Fig. 2, the frame rate is shown. Here the image is being re-gridded and refreshed at 14.8 fps on a a Macbook Pro with an Nvidia GeForce GT 750M. On a Xeon workstation with an Nvidia GeForce TITAN X, the rate hit the cap of 60 fps.

Discussion

The raw data size coupled with the number of reconstruction parameters greatly increases the effective dimensionality of the image data. The raw data was 660 MB, but with over 20,000 potential slice positions (one per profile), ~100 different useful slice thicknesses, ~10 kernel radii, ~5 oversampling factors, and 2 coil combination methods, one could create roughly 200 million different 384 x 384 images from this tool. Creating and storing the roughly 236 TB of information they would contain would be ridiculous, but using CITRON the same amount of information is responsively available to the fingertips of a radiologist while using only 660 MB on disk.

For archival purposes, creating images on demand from the raw data means that information collected at scan time is never lost. All information collected at the scanner is retained, so images can be improved with time, as better reconstruction techniques emerge, or used to answer different clinical questions.

One example use case would be to distinguish between a pathology and an imaging artifact. With the ability to vary the reconstruction parameters, the user can either minimize the artifacts or watch them change to see whether the suspicious feature was persistent or varied like an artifact.

Another use case would be to shift slice positions to avoid volume averaging. With CITRON a few key presses is sufficient to center a feature in a slice if it happened to cross slice boundaries upon initial reconstruction.

Conclusions

We developed a tool called CITRON to interactively explore MRI data in real time and thus avoid the need to reconstruct a single static image set and throw away the original scan data. CITRON could be used to better answer clinical and research questions by allowing users to control a live reconstruction during viewing.

Acknowledgements

NCI K25 CA176219, NCATS UL1 TR000445

References

[1] http://github.com/davidssmith/CITRON

[2] Sengupta et al. 2015, MRM, Early View (doi: 10.1002/mrm.25848)

[3] http://github.com/davidssmith/TRON

[4] Walsh et al. 2000, MRM, 43, 682.

Figures

Figure 1: Diagram of CITRON subsystems. The OpenGL front end handles the image display and user input and sends parameters to the reconstructor back end. The reconstructor sends the updated image back to the OpenGL subsystem for refreshing the display.

Figure 2: Screenshot of the CITRON window, showing a live, interactive reconstruction of continuous golden angle radial data with the ability to instantly vary reconstruction parameters, including slice position and thickness down to 70 microns.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1782