“If health systems do not provide institutional access to [foundation models], they implicitly tolerate entirely unsupervised use.”
— Dr. Vera Sorin et al., European Radiology 2026
Welcome to Radiology Access! Your biweekly newsletter on the people, research, and technology transforming global imaging access.
In this issue we cover:
RSNA AI Theater w/ Campbell Arnold
Beyond T1 and T2: Teaching AI to Understand MRI Contrast
Can NVIDIA’s New 3D Synthetic Imaging Generator Improve Your AI Algorithm’s Performance?
The Canary in the AI Mine: Post-deployment Monitoring for Foundation Models
If you want to stay up-to-date with the latest in Radiology and AI, then don't forget to subscribe!
RSNA AI Theater w/ Campbell Arnold

I'm excited to share that I'll be presenting on behalf of Subtle Medical at this year's RSNA AI Theater in Chicago!
Join me on Sunday, November 29, from 1:30–1:50 PM at the AI Theater (Booth #5536) for Beyond Better Images: The Power of AI Across MR, PET, and CT. I'll discuss how Subtle Medical’s AI-powered image enhancement is evolving beyond simple image quality improvements to help organizations get more from their existing scanner fleets. We'll explore applications across MR, PET, and CT, including opportunities to improve clinical workflows, increase efficiency, and extend the value of existing imaging infrastructure.
If you're attending RSNA, I'd love to see you there!
Beyond T1 and T2: Teaching AI to Understand MRI Contrast
Could DICOM metadata help AI better understand the images it's processing?

You can recognize a T1-weighted scan at a glance, but how would you teach an AI algorithm what makes it a T1-weighted image? Two exams can look quite different depending on the scanner, acquisition parameters, and imaging protocol used to produce them. This becomes particularly challenging when working with large imaging datasets, where sequence labels and DICOM metadata may be missing, inconsistent, or unreliable. A new paper introduces MR-CLIP, a model designed to learn the relationship between MRI appearance and acquisition parameters without relying on manually assigned sequence labels.
The approach adapts the contrastive learning concept behind CLIP to align MR images with their corresponding DICOM metadata in a shared representation space. Rather than simply classifying images as T1 or T2, the model learns more detailed contrast representations using acquisition parameters such as TE, TR, and TI, alongside scanner and sequence information. The authors trained the model on an impressive dataset of 169,634 brain MRI volumes from over 40,000 subjects, consolidating more than 21,000 unique acquisition configurations into 1,415 contrast-aware groups.
MR-CLIP achieved a 78.7% top-1 retrieval rate for 3D scan-to-metadata retrieval, compared with 67.4% for another state-of-the-art model, BiomedCLIP. Performance was particularly strong when retrieving images from metadata descriptions, achieving 90.9% top-1 retrieval compared with 38.5% for BiomedCLIP. The authors also demonstrated promising generalization using the independent OASIS dataset, even though some metadata fields were unavailable.
What's particularly interesting is the potential beyond sequence identification. More robust representations of MRI contrast could help with image retrieval, cross-site data harmonization, and potentially image synthesis. The current work is limited to brain MRI, and further validation is needed, but it offers an interesting approach to making better use of information already embedded in clinical imaging archives.
Both the open source code and the Machine Learning in Medical Imaging article are available online.
Bottom line: MR-CLIP demonstrates how AI can use acquisition parameters to learn richer representations, potentially helping make heterogeneous imaging datasets easier to organize.
Can NVIDIA’s New 3D Synthetic Imaging Generator Improve Your AI Algorithm’s Performance?
Can faster, controlled synthetic imaging help solve medical imaging’s training data bottleneck?

Synthetic medical images are only useful if they get the important details right. A generated CT scan may look realistic, but if a small tumor is missing or appears in the wrong location, it makes for a poor training example. At the same time, synthesizing high-resolution 3D volumes can be computationally expensive. A new paper presented at AAAI 2026 introduces MAISI-v2, a framework from NVIDIA designed to address both challenges, accelerating 3D image generation and improving how closely synthetic images adhere to anatomical expectations.
To accelerate generation, the authors replace the traditional denoising process used in diffusion models with rectified flow. Rather than gradually removing noise through thousands of iterative steps, rectified flow learns a more direct path from noise to an image, substantially reducing the number of steps required (from 1,000 to just 30). For a 1 mm isotropic 512 × 512 × 512 CT volume, MAISI-v2 reduced total generation time from approximately 213 seconds to 21 seconds, without sacrificing image quality.
The second contribution focuses on image fidelity. The authors introduce a region-specific contrastive loss that encourages the model to preserve changes in a targeted region while keeping the rest of the image consistent. For example, the model generates two images from the same noise input, one conditioned on the original tumor segmentation mask and another with the tumor label replaced by surrounding tissue. Training encourages the outputs to differ in the tumor region without unnecessarily changing the surrounding anatomy.
However, can better synthetic images actually make better AI models? The authors put this to the test by evaluating five downstream segmentation tasks. Adding synthetic data to real training datasets improved Dice scores with gains of 3.3 to 7.5 percent compared with training on real images alone. The code is available on github.
Bottom line: NVIDIA’s MAISI-v2 demonstrates how advances in generative AI can make realistic, high-resolution 3D medical image synthesis that actually improve downstream segmentation performance.
The Canary in the AI Mine
Why traditional post-deployment monitoring approaches aren’t sufficient for foundation models.

Traditional AI monitoring has focused on tracking performance, model drift, or input changes. However, foundation models introduce new challenges to post-deployment monitoring because their behavior depends not only on model weights, but also on prompts, retrieved information, tool permissions, and deployment settings. Additionally, many are stochastic rather than deterministic, meaning identical inputs and configurations can produce different responses. This makes it harder to reproduce and retrospectively evaluate a specific interaction.
A new opinion piece in European Radiology argues that reviewing generated responses alone is no longer sufficient, health systems now need observability. This means the ability to reconstruct which inputs, settings, information sources, and system pathways shaped a foundation model’s output. Institutionally managed deployments are essential to provide the configuration control, visibility, and audit trails necessary for meaningful deployment oversight.
The authors also recommend prospective canary testing by routinely running a rotating set of scenarios against production systems to detect emerging failures. This can be complemented by monitoring unexpected tool use, drops in retrieval confidence, model disagreements, and inconsistencies between generated narratives and underlying images or source data.
Bottom line: Foundation model post-deployment monitoring must extend beyond simple output based evaluations, including tracking context like prompts and retrieved documents.
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
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.



