- Campbell Arnold
- Feb 25
- 7 min read
Have No Fear, SubtleHD is Here!
This is a big moment for Subtle Medical—one that I’m proud to have been involved in. We’ve just received FDA clearance for SubtleHD, our fourth clearance in the past year! SubtleHD is the most advanced MR acceleration algorithm we’ve developed thus far and rounds out our new ELITE package—three algorithms that work in concert to accelerate MRI protocols, improve image quality, and streamline workflows.
SubtleHD: Performs combined denoising and sharpening to enhance accelerated scan times up to 80%. Works on all anatomies and common sequences.
SubtleSYNTH: Generates a synthetic spine STIR from acquired T1 and T2 sequences, providing additional spine protocol time savings beyond enhancement acceleration.
SubtleALIGN: Automatically aligns 3D brain images for improved consistency across sites and for longitudinal studies.
Having worked on the development and regulatory clearance of these tools, I know firsthand how transformative they can be for radiology. Their vendor-neutral design allows this technology to be compatible with almost all MRI scanners, allowing for a single solution to work across large institutions with multiple different magnets. These acceleration and workflow tools can enhance imaging center efficiency, increase the patient population a single magnet can treat, and extend scanner lifespans—reducing the need for large capital investments.
For me, this clearance represents more than just another milestone—it’s a testament to the power of AI to shape the future of medical imaging and I can’t wait to see the impact it has in real-world clinical settings.
Foundation Models are Knocking at the FDA’s Door
In regulatory news, Aidoc has secured FDA clearance for the first device powered by a foundation model, marking a significant regulatory milestone. Unlike traditional AI models which are designed for single-use cases, foundation models are built for broader applicability, enabling faster development and potentially improving generalization and performance. As expected, rather than granting broad clearance for foundation models, the FDA is approving them only for specific use cases—in this case, Aidoc’s CARE1 model is part of their CT rib fracture triage solution.
This milestone not only highlights the growing role of foundation models in healthcare but also provides insight into how the FDA is approaching their regulation, at least for now. With ongoing discussions at the agency regarding the oversight of generative AI, the regulatory landscape may continue to evolve. Meanwhile, Aidoc has already teased the development of CARE2, its next-generation foundation model, with plans to expand into additional clinical applications.
But is Anyone Home at the FDA?
The past two weeks have been turbulent for the FDA, with widespread staff cuts around Valentine’s Day, followed by reinstatement emails sent to some employees last Friday. The Center for Devices and Radiological Health (CDRH), which oversees AI-enabled imaging devices, was hit particularly hard, though it remains unclear how many regulators have now returned. The reinstatements came after pushback from the medical device industry, but the episode underscores the fragile state of AI regulation as the FDA grapples simultaneously with staffing changes and with updating their frameworks to accommodate generative AI, continuously learning algorithms, and foundation models.
With over 1,000 AI-based medical devices cleared, the demands on the FDA have been rapidly increasing. To improve the regulatory process for AI-driven technologies, the agency has been developing new initiatives, including post-market surveillance guidance and policies for managing predetermined device changes. However, the recent staffing instability and shifting policy priorities create uncertainty for both regulators and developers. Will these disruptions slow the pace of AI approvals, and will the FDA continue evolving its frameworks to better accommodate the next generation of medical AI?
ASSR 2025 Recap
The American Society of Spine Radiology (ASSR) 32nd Annual Conference wrapped up in San Diego last week. It was an exciting meeting filled with cutting-edge research, thought-provoking discussions, and a glimpse into the future of spine imaging. Here are some key takeaways from the conference:
AI Continues to Improve Low-Field Image Quality: Amish Doshi gave an excellent presentation on low-field MRI for kinematic spine imaging. His examples highlighted how flexion and extension imaging can illuminate pathology that’s easily missed in standard scans. Additionally, he showcased that low-field image quality continues to improve with the advent of deep learning based image enhancement algorithms.
AI Copilots—Assistance, Not Autonomy: Adam Flanders gave a whirlwind overview of AI copilots and their future role in radiology. With imaging volumes increasing and vastly outpacing trainees entering the field, radiologists are looking for ways to improve efficiency. AI-driven assistance is being seen more as necessary a means to lighten workloads. The key message though was that Radiologists are still in the pilot’s seat—AI is a copilot, and ultimate responsibility rests with the physician.
Spine Protocol Optimization and Gadolinium Dose Reduction: It was a privilege to share two of our latest projects in the Best Diagnostic Papers session. I spoke about accelerating spine protocols using combined image enhancement and image synthesis. A timely talk given that our enhancement algorithm, SubtleHD, was cleared by the FDA the day before! The second talk covered our preliminary work developing a gadolinium contrast dose reduction algorithm for the spine, which could reduce contrast exposure, costs, and environmental impacts.
Photon-Counting CT (PCCT) is Here—But So Are Its Challenges: Tim Amrhein delivered an impressive talk on PCCT, showcasing its advantages in reducing radiation dose while increasing resolution. The images were stunning, but there are still hurdles to overcome. With better resolution comes larger data sizes, and transferring these massive datasets from the scanner can take up to an hour—posing workflow challenges that need to be addressed.
Deep Learning Accelerated MRI—The Future is Now: Larry Tanenbaum presented exciting applications of deep learning for MRI acceleration. His talk covered algorithms from a wide variety of vendors, including three from Subtle Medical: SubtleHD for image enhancement, SubtleSYNTH for STIR sequence synthesis, and AiMIFY developed in collaboration with Bracco for gadolinium contrast boosting. These tools are pushing the boundaries of MRI efficiency and image quality.
Synthetic CT—Reducing Radiation, Expanding Possibilities: Johan Van Goethem gave an excellent presentation on synthetic CT generation using MRI, powered by MRIguidance’s BoneMRI. This technique is already making a clinical impact, particularly in reducing radiation exposure for pediatric patients. One particularly compelling case showcased how BoneMRI helped avoid fetal radiation exposure during pregnancy.
Challenging the Status Quo in Radiology: Wende Gibbs gave a standout talk near the conference close which challenged radiologists to rethink their practice. She asked two tough but necessary questions: 1) If we focused on patient outcomes rather than scan volume, how would our workflows change? 2) Are our reports truly serving patients and referring physicians? It was a powerful reminder to keep patient impact at the forefront of radiology.
ASSR 2025 was a fantastic meeting, and I left inspired by the innovative research, the insightful discussions, and the commitment to excellence of the spine radiology community. As spine imaging continues to evolve, it’s clear that deep learning and AI will play an increasingly important role in shaping the future of the field.
TotalSegmentator MRI: Multi-anatomy, Multi-sequence, Multi-modal!
Multi-anatomy multi-sequence segmentation is having a moment, it seems like every issue of RadAccess features a new algorithm! In this edition, we’re covering the recent Radiology publication on TotalSegmentator. One unique aspect of TotalSegmentator is that it’s also multi-modal, working on both CT and MR images. The original algorithm was CT only and we covered the v2 release in 2023. Last year, the authors announced TotalSegmentator now supports MR on arXiv. This latest Radiology article appears to extend TotalSegmentator even further, now covering 80 total anatomical structures and more than doubling the training dataset. This advancement builds upon the success of TotalSegmentator CT and aims to develop robust and generalizable multimodal segmentation.
TotalSegmentator was trained on 1,143 scans (616 MRI, 527 CT) and demonstrated high segmentation accuracy, achieving a Dice score of 0.839 on an internal test set. It outperformed existing MRI segmentation models, including MRSegmentator and AMOS, and performed comparably to TotalSegmentator CT on a separate CT test set. The model was also applied to a large dataset of over 8,000 abdominal MRI scans and confirmed expected organ volume changes with age. Importantly, the integration of both MRI and CT data into training improved MRI segmentation performance, highlighting the benefits of multimodal learning.
By offering a fully automated, sequence-independent MRI segmentation tool, TotalSegmentator MRI has the potential to enable better clinical integration of segmentation and quantitative analysis into radiology workflows. With its code, dataset, and an easy-to-use online segmentation tool openly available, this work represents a major step toward more efficient, reproducible, and accessible MRI analysis across clinical and research settings.
First-In-Patient Scan for New Low-field Device
Exciting times for Wellumio, a New Zealand-based low-field device start-up. They scanned their first clinical patient, which marks the beginning of a clinical study evaluating their Axana 0.1T portable MRI for stroke detection. The portable device is a head-only scanner and has a compact cart design. The study is being conducted in partnership with the Australian Stroke Alliance and will assess the feasibility, safety, and usability of Axana in both healthy individuals and stroke patients. Researchers will also compare its diffusion-weighted imaging (DWI) capabilities to standard hospital MRI. The hope is that Axana will streamline stroke triage, allowing for informed decisions and rapid intervention during these emergent cases.
Resource Highlight
Looking for more radiology news? Check out The Imaging Wire, it’s one of my favorite ways to stay updated on industry trends. This publication posts twice a week, offering an in-depth exploration of a single topic while also providing a brief overview of other notable developments. Whether you're a first year graduate student or seasoned professional, The Imaging Wire is a fantastic resource to keep you informed on the latest in medical imaging.
Feedback
We’re eager to hear your thoughts as we continue to refine and improve RadAccess. Is there an article you expected to see but didn’t? Have suggestions for making the newsletter even better? Let us know! Reach out via email, LinkedIn, or X—we’d love to hear from you.
References
https://www.aidoc.com/about/news/aidoc-secures-landmark-fda-clearance/
https://theimagingwire.com/2025/02/19/staff-cuts-have-occurred-at-the-fda/
Akinci D’Antonoli, Tugba, et al. "TotalSegmentator MRI: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in MRI." Radiology 314.2 (2025): e241613.
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.