“AI is still peripheral. It only touches our workflow at the edges.”
— Dr. Paul Chang, Vice Chair of Radiology at University of Chicago and creator of Stentor
Welcome to Radiology Access! Your biweekly newsletter on the people, research, and technology transforming global imaging access.
In this issue we cover:
TotalSpineSeg: Building a Robust Spine Segmentation Algorithm
Better Images, Better AI: An Overlooked Role for Image Enhancement
The Last Mile of Radiology AI: You’ve Got the Algorithm, But Are You Getting the Results?
FREE Resource: EuSoMII Webinar Series
If you want to stay up-to-date with the latest in Radiology and AI, then don't forget to subscribe!
TotalSpineSeg: Building a Robust Spine Segmentation Algorithm
How anatomical knowledge and thoughtful model design can overcome the challenges of real-world imaging variability.
Developing a clinically useful spine segmentation algorithm is surprisingly challenging. Most existing approaches focus on a single image contrast, provide only coarse anatomical labels (e.g., “lumbar vertebrae” rather than “L2”), and struggle with the wide variation in acquisition parameters and image quality encountered in clinical practice. Compounding the problem, publicly available spine MRI datasets are limited, and many lack consistent anatomical labels.
A recent European Radiology article aims to address these challenges with TotalSpineSeg, a publicly available algorithm designed to provide individual vertebrae and disc-level segmentations across any MRI sequence. The authors trained the algorithm using more than 1,400 publicly available MRIs and evaluated it on over 500 additional cases spanning different sequences, sites, modalities, and pathologies.
The approach uses a two-stage cascaded model. First, it identifies the intervertebral discs; then, it labels vertebrae and key anatomical landmarks across the cervical, thoracic, lumbar, and sacral spine. These landmarks help assign individual anatomical labels to each vertebra and disc (e.g., L5). The authors also developed an option to incorporate a localizer scan, improving anatomical labeling of small-field-of-view axial scans that may not include the landmarks needed for reliable level identification.
Building a robust model required an equally thoughtful training strategy. The authors leveraged a broad range of data augmentation techniques, including contrast manipulation, artifact simulation, spatial transformations, intensity redistribution, and intensity inversion. The segmentation network itself is based on a 3D nnU-Net architecture, with the paper detailing the authors' exploration of different training configurations.
The results were impressive. TotalSpineSeg demonstrated robust segmentation of four anatomical structures (vertebrae, intervertebral discs, spinal cord, and spinal canal) and reported greater than 99% accuracy in anatomical structure labeling, outperforming other state-of-the-art methods. Perhaps more importantly, the algorithm generalized to previously unseen MRI contrasts, including PSIR, STIR, and MP2RAGE, and even demonstrated applicability to CT.
This study highlights that clinical robustness requires designing algorithms around the variability of real-world imaging, not just the consistency of curated training datasets. The extensive augmentation strategy helped expose the model to variations in contrast, artifacts, acquisition parameters, and image quality. Just as importantly, the authors' understanding of spinal anatomy informed the model architecture itself. Their cascaded approach uses disc identification to guide vertebral labeling, helping accommodate anatomical variants such as patients with four or six lumbar vertebrae. Models without this two-stage strategy performed substantially worse, underscoring how clinical insight can, and should, influence algorithm design.
Importantly, TotalSpineSeg is publicly available and has been integrated into the Spinal Cord Toolbox, making it accessible for further research and potential clinical applications.
Bottom line: TotalSpineSeg demonstrates how combining anatomical expertise and thoughtful model design can produce robust algorithms that generalize across real-world variability.
Better Images, Better AI: An Overlooked Role for Image Enhancement
Could enhancing MRI scans before training detection algorithms improve their diagnostic accuracy?
We all know the scene from CSI: Miami: they've got the perp, but all they have is a blurry CCTV image of the license plate. A few dramatic keystrokes and someone yelling “Enhance that!” later, the image magically resolves into WYW-1747.
MR image enhancement isn't quite that dramatic, but a new study suggests it can do more than just improve image quality on accelerated scan. It may also make training downstream diagnostic AI models more accurate, even when the original images are already diagnostic-quality.
In the Journal of Magnetic Resonance Imaging, researchers applied SubtleHD(MR) enhancement to T1-weighted brain MRI scans from Alzheimer's Disease Neuroimaging Initiative dataset. They then trained deep learning models to classify patients as cognitively normal, having mild cognitive impairment, or having Alzheimer's disease.
Training and testing on enhanced images increased accuracy from 85.2% to 88.7% and macro-AUC from 0.951 to 0.968. The strongest results came when enhanced images were used for both training and inference, but models trained on enhanced images also outperformed standard-trained models when evaluated on the original, unenhanced scans. Applying enhancement only at inference, however, did not improve a model trained on standard images. The findings were also replicated in an external dataset from the National Alzheimer's Coordinating Center, with an even greater performance difference (macro-AUC = 0.679 standard vs 0.819 enhanced).
What makes this study particularly interesting is that they ran enhancement algorithms on standard-acquisition images that were already considered diagnostic-quality, not accelerated or degraded scans. This means this enhancement benefit could potentially be applied retrospectively to other existing public datasets or private imaging archives before training diagnostic AI models to improve their final accuracy.
We typically think of image enhancement as a way to accelerate scans, but this study highlights the potential for preprocessing imaging data to improve downstream AI model accuracy.
Bottom line: Enhancement algorithms may have a role beyond scan acceleration. Using them to preprocess images before diagnostic AI model training could improve detection accuracy.
The Last Mile of Radiology AI: You’ve Got the Algorithm, But Are You Getting the Results?
Why workflow integration, not just algorithm speed, determines whether AI delivers clinical value.
What good is an AI algorithm if its results arrive after the radiologist has signed the report? A recent study in the Journal of the American College of Radiology takes a closer look at this question, evaluating 4.5 years of AI deployment across a 20-center outpatient radiology network. The team analyzed nearly 100,000 AI-assisted examinations, spanning 10 applications from 7 vendors. They investigated infrastructure latency, reporting efficiency, and radiologist adoption.
One of the most interesting findings was that 72% of total AI latency was attributable to data routing rather than algorithm processing. Median end-to-end latency was approximately 2 minutes (IQR: 1.74–3.05 minutes), with the time required to retrieve, upload, and download imaging studies accounting for most of the delay. Across over 20,000 exams, 7.2% of AI results arrived after report finalization. This varied considerably by application, reaching 13.2% for chest CT compared to just 3.0% for knee MRI.
The authors also found measurable improvements in reporting efficiency. AI availability was associated with a 26% reduction in report turnaround time for trauma radiography and an 18% reduction for knee MRI. However, gains were not universal. MRI brain volumetry showed no statistically significant change in turnaround time.
Surprisingly there was a disconnect between measured efficiency and radiologist perception. Despite the turnaround-time improvements, perceived productivity remained below the neutral midpoint. The authors suggest that time saved through AI may simply be absorbed by additional case volume, rather than translating into a lighter workload.
The study offers a valuable real-world look at what happens after AI leaves the validation environment and enters clinical practice. For AI developers, it illustrates why deployment pipeline efficiency may matter more than inference speed. The authors advocate defining time-to-display targets before deployment and continuously monitoring latency, algorithm performance, and user satisfaction. For radiologists, the findings reinforce that the value of AI depends not only on what it can do, but when and how it enters their workflow.
Bottom line: AI’s clinical value depends not just on algorithm performance, but on delivering results quickly and reliably into the radiologist’s workflow.
FREE Resource: EuSoMII Webinar Series
The European Society of Medical Imaging Informatics (EuSoMII) hosts a free webinar series. It’s an excellent opportunity to hear emerging developments in imaging informatics from leading experts in the field. On Nov 12th at 12PM EST their featured talk covers the JACR article we highlighted in this week’s RadAccess issue! The EuSoMII webinar series is a useful resource for radiologists, researchers, and AI developers looking to stay current with developments in the field of imaging informatics.
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.
