2nd Edition | Rocket Ships & Accelerated Sequences

Low-Field MRI and the Final Frontier

Since the advent of portable low-field scanners, they’ve been showing up in the strangest places! Antarctica, the Motorcycle Grand Prix, in the back of vans, where will they show up next? Well you can add launchpads to the list. Our feature article this week covers the recent events involving portable MRI and rocket ships.

One giant leap for scankind

Photo credit: Polaris Program / John Kraus

MRI took one step closer to becoming an interstellar imaging modality this month. While a scanner hasn’t been launched into space yet, portable MR systems are getting closer to the launchpad. In a recent Hyperfine press release, the company announced a collaboration with Medical University of South Carolina researcher Donna Roberts to study brain health in astronauts during the SpaceX Polaris Dawn mission.

The crew was launched into space September 10th and landed back on Earth September 15th. Sources confirmed the astronauts were successfully scanned 3 times: prior to launch, within hours of landing, and one day after return. Researchers hope to use these images to study spaceflight-associated neuro-ocular syndrome (SANS). The portable MRI allowed researchers to collect brain scans mere hours after re-entry, which has not previously been done. This will help researchers distinguish spaceflight associate brain changes from those associated with re-adaption of the astronauts’ brains to Earth’s gravity.

Study lead, Donna Roberts stated “This study represents the first step towards a future where advanced medical imaging such as MRI is available in space, such as on a future space station, lunar base, or spaceship supporting human missions to Mars.” We’ll be looking out for the upcoming study results, and keeping our eye on the international space station for a future scanner.

Accelerating terrestrial imaging

While portable MR is bringing imaging closer to the stars, there is no shortage of need for additional imaging capacity here on Earth. While much of the hype surrounds AI and medical imaging initially focused on diagnostic algorithms, the past few years have seen significant gains in deep learning based image reconstruction, both for accelerating scan times and improving image quality. More recently, image synthesis has offered an additional method for accelerating scan protocols by eliminating the need to collect entire sequences. Four recent studies explored AI applications that could improve scanning efficiency.

Speedy sequences for scanning the whole body in 60 minutes

Recently, whole-body MRI companies have made big headlines ranging from celebrity endorsements to scathing editorials from skeptics. Whole-body MRI companies like Ezra and Prenuvo offer direct-to-consumer preventative screening MRI for approximately $2,000-2,500. While these scans are generally marketed for early cancer detection, many in the radiology community remain apprehensive, as these services are offered to consumers outside recommended screening criteria.

We’re beginning to gain more insight into how these companies can offer whole-body screening in a 60 minute scan session. Ezra recently published a clinical validation of their accelerated MR denoising algorithm for brains in Radiology Advances. Their approach uses parallel imaging to accelerate scans and restores image quality by applying a UNet trained to remove simulated noise and artifacts from degraded images. Approximately 2000 subjects were used for model training and validation, and roughly 100 (varies by test) were included in the quantitative testing and qualitative review by 5 radiologists. The researchers showed that despite reducing scan times by 29%, they were about maintain or exceed standard-of-care quality across qualitative metrics (e.g. image quality, pathology visibility, etc.) and quantitative metrics (i.e. SNR, CNT, FWHM). These time savings will help Ezra squeeze five anatomical regions into a 60 minute session.

Synthesized sequences beget more synthesized sequences

Researchers at the Technical University of Munich study published an interesting two staged image synthesis approach in European Radiology. While many studies have explored synthesizing sequences from acquired scans, this team developed two cascading networks with the second model depending on the first network’s synthesized output.

The first model generates a sagittal T1w from acquired sagittal T2w and axial T1w Dixon sequences. The second algorithm then produces a sagittal STIR from an acquired T2w and the previously synthesized sagittal T1w. Both networks used the 3D Pix2Pix architecture and were trained using over 3000 and 150 subjects, respectively. Quantitative tests showed improved image quality relative to networks trained using only the T2w image as an input. Images were also reviewed by 7 radiologists in a turing test, which resulted in a misclassification rate around 40%, meaning readers had a difficult time distinguishing real and synthetic images. While further clinical validation is still needed, this could further accelerate spine protocols and permit additional sequences they are usually not collected due to time constraints.

It will be interesting to see how radiologists & researchers react to synthesized images that depend on synthesized inputs. A recent Nature article raised alarms about AI model collapse when training with recursively generated data. While the publication was about text-based models, many researchers are likely to have similar concerns about recursive image generation.

Accelerated sequences squeeze vessel wall imaging into protocols

Intracranial vessel wall MRI has proven more effective than conventional vessel imaging techniques for differentiating vascular diseases. However, vessel wall sequences typically have long scan times and low SNR due to their high resolution and flow suppression. In a recent AJNR article, researchers at the University of Washington and Subtle Medical (full disclosure, my employer) used a deep-learning-based image denoising algorithm to accelerate vessel wall sequences and improve image quality.

The researchers compared a traditional vessel suppressed T1 sequence, an accelerated flow suppressed sequence (36% faster), and the accelerated sequence after processing with a denoising algorithm. Two readers reviewed the images and rated them on a variety of qualitative measures (e.g. image quality, SNR, flow artifacts, etc.). As expected, readers found the flow suppressed sequences had less artifacts than the traditional vessel suppressed T1. Additionally, readers found denoised accelerated sequences to have improved SNR over accelerated sequences. A quantitative analysis of vessel wall SNR and CNR similarly showed improvements after denoising. Overall, the scan time reduction, reduced flow artifacts, and SNR improvement combine to make a compelling argument for squeezing these sequences in protocols.

Accelerating cardiac imaging 20x with CineVN

In a final acceleration study, Siemens researchers and academic collaborators developed CineVN, a deep-learning-based method for reconstructing highly accelerated cardiac cine MRI scans with high spatiotemporal resolution. Their method, presented in Magnetic Resonance Medicine, expands on a Variational Network (the VN in CineVN) to improve data consistency and image quality for spatiotemporal data.

CineVN was tested on both retrospectively undersampled images, which were accelerated up to 20x, and prospectively accelerated data (8-16x) collected over 1-2 heartbeats per slice. Their model handily outperformed two other state-of-the-art reconstruction techniques in terms of image quality and provided more accurate measurements of ventricular volumes and myocardial strain. This accelerated approach has the potential to expand cardiac imaging access, particularly for patients with arrhythmia or those who struggle with breath-hold requirements.

Trading CT for MR

CT has long been a mainstay for emergency imaging, but scan volumes are continually increasing each CT exposed patients to radiation. Two recent studies explored advancements in MRI that could challenge the dominance of CT, reducing MR scan time and increasing resolution. These developments point to safer, faster alternatives that could reduce radiation exposure and ease the burden on radiology services.

A 3 minute high-resolution CT-like MR sequence

Head CT has long enjoyed its status as standard-of-care for craniofacial imaging thanks to the high spatial resolution, excellent bone contrast, and short scan times the modality offers. However, researchers at the University of Pennsylvania recently proposed a MR competitor that offers comparable resolution and scan time, without the long-term risks associated with CT radiation.

In this Magnetic Resonance in Medicine article, the group demonstrated their ultrashort echo time (UTE) sequence on 10 adult volunteers and 3 pediatric patients. The UTE sequence had similar scan time and resolution to a traditional head CT, requiring only 3 minutes scan time (~30s for a CT) and having 0.65mm isotropic voxels (~0.5-0.75mm for CT). Quantitative comparisons and craniometric measurements made on skull segmentations showed strong similarity between head CT and UTE sequences. While this is early evidence, the method shows promise for lower radiation exposure in patients that would otherwise receive repeat head CTs.

A 7.5 minute MR protocol to replace CT angiography

Head and neck CT angiography are increasingly being ordered for patients with nonfocal neurological symptoms, which has resulted in a strain on radiology services. Additionally, as mentioned in the previous article, head CT is associated with long-terms risks from radiation exposure. In this AJNR article, researchers Stanford and GE Healthcare evaluated a fast 7.5 minute MRI protocol that could supplant head and neck CT angiography.

The study retrospectively compared over 170 cases with CT angiography and a follow up MRI within 24 hours. The 7.5 minute protocol included GE’s 2.5 minute multicontrast sequence (NeuroMix) and a 5 minute magnetic resonance angiography (MRA) sequence. Four radiologists reviewed the images and determined consistency of results between modalities. They found that NeuroMix + MRA demonstrated equivalent or better results in 95% of cases, with the remaining 5% of cases having missed pathology due to the smaller FOV in MR images. These results suggest that MR could be a radiation-free alternative to CT angiography for patients with acute neurologic presentations. Additionally, collecting only MR imaging as opposed to CT and later an MR follow up could significantly reduce radiology workloads and demand on imaging services.

Resource Highlight

Looking for more radiology news? Check out The Radiology Report Podcast, it’s an excellent long-form podcast that interviews radiology leaders and industry experts. The conversations are candid and offer both career advice and in-depth explorations of radiology topics. They’re always a pleasure to listen to and give you great insights into how leaders in radiology think. Whether you're a first year graduate student or seasoned professional, The Radiology Report Podcast is a fantastic resource to keep you informed and entertained.

Feedback

As a new resource, we're all ears and eager to know your thoughts on how to improve our newsletter. Don't see an article you thought should be included? Send us an email, or reach out to use via X.

References

Hyperfine Press Release: Researchers to Monitor Astronaut Brain Health Using Images Acquired with the Hyperfine, Inc. Swoop® Portable MR Imaging® System. https://hyperfine.io/about/news/press-release-researchers-to-monitor-astronaut-brain-health-using-images-acquired-with-the-hyperfine-inc-swoop-portable-mr-imaging-system
Onac, Laura, et al. "An image-domain deep-learning denoising technique for accelerated parallel brain MRI: prospective clinical evaluation." Radiology Advances 1.3 (2024): umae022.
Graf, Robert, et al. "Generating synthetic high-resolution spinal STIR and T1w images from T2w FSE and low-resolution axial Dixon." European Radiology (2024): 1-11.
Kharaji, Mona, et al. "DANTE-CAIPI Accelerated Contrast-Enhanced 3D T1: Deep learning-based image quality improvement for Vessel Wall MR." American Journal of Neuroradiology (2024).
Vornehm, Marc, et al. "CineVN: Variational network reconstruction for rapid functional cardiac cine MRI." Magnetic Resonance in Medicine (2024).
Vu, Brian‐Tinh Duc, et al. "Three contrasts in 3 min: Rapid, high‐resolution, and bone‐selective UTE MRI for craniofacial imaging with automated deep‐learning skull segmentation." Magnetic Resonance in Medicine.
Decker, Johannes H., et al. "NeuroMix with MRA: A Fast MR Protocol to Reduce Head and Neck CTA for Patients with Acute Neurologic Presentations." (2024).

Disclaimer: There are no paid sponsors of this content. The opinions expressed are solely those of the newsletter authors, and do not necessarily reflect those of referenced works or companies.