The research, published in PLOS Digital Health by a team led by Shaokai Zheng and David Bervini, focused on so-called vulnerable plaques in carotid atherosclerosis. These are lesions prone to rupture that contain features such as intra-plaque hemorrhage, a lipid-rich necrotic core, or a ruptured fibrous cap. Once a plaque ruptures, thrombo-embolic debris can travel to the brain and cause a stroke. Detecting such plaques matters because their prevalence among patients with asymptomatic carotid stenosis is estimated at around 26 percent, and identifying them guides decisions about carotid endarterectomy, the surgical removal of plaque from the artery.

Today, the preferred way to assess plaque vulnerability is multi-contrast plaque MRI, which can visualize hemorrhage and cap integrity directly. But MRI scanners with the right sequences and the radiological expertise to interpret them are concentrated in tertiary centers, creating bottlenecks and diagnostic delays. The Bern team asked a different question: could the physics of blood flow itself, simulated on ordinary clinical data, reveal the fingerprint of a dangerous plaque?

Their workflow began with nine patients scheduled for carotid endarterectomy between January 2021 and March 2023, yielding eighteen carotid arteries for analysis, nine of which were rated vulnerable on plaque MRI according to the Plaque-RADS grading system. From each patient’s CT angiogram, the researchers segmented the carotid bifurcation, the fork where the common carotid artery splits into the internal and external branches, and reconstructed three-dimensional geometries containing between 1.3 and 1.8 million computational cells. Crucially, instead of imposing generic pulsatile waveforms, they anchored the simulations in reality by using peak systolic and diastolic flow velocities measured with neurovascular ultrasound, a standard first-line examination, to set the boundary conditions of the model.

The team then solved the Navier-Stokes equations for blood flow, modeled as a Newtonian fluid under laminar conditions, at both peak systole and diastole using steady-state simulations in COMSOL Multiphysics. Each simulation took under an hour on an ordinary desktop computer. From the resulting velocity fields, they extracted an unusually rich set of 289 features, including geometric measures such as bifurcation angles and stenosis ratios, and hemodynamic quantities such as wall shear stress, vorticity, and helicity, a measure of the corkscrew rotation of flow. Importantly, they went beyond simple averages, computing higher-order statistics such as skewness and kurtosis, which describe the asymmetry and tailedness of how these quantities are distributed inside the vessel.

Two machine learning classifiers, logistic regression and a linear support vector machine, were trained to distinguish vulnerable from non-vulnerable arteries. To guard against overfitting in this tiny dataset, the researchers used leave-one-out cross-validation with recursive feature elimination embedded inside each fold, preventing any data leakage, alongside correlation filtering and L2 regularization. The classifiers achieved mean areas under the receiver operating characteristic curve of 0.77 and 0.78 respectively, moderate discriminative performance with wide confidence intervals that reflect the small sample size.

The most striking result was which features the algorithms kept selecting. Mean helicity emerged as the strongest protective signal, with an odds ratio of 0.5, meaning more organized helical flow was associated with non-vulnerable plaques. Higher kurtosis of helicity, indicating concentrated and coherent helical flow structures, was likewise protective, while skewness of streamwise vorticity, which captures how swirling structures aligned with the vessel axis are distributed, ranked among the strongest risk predictors. Eight features were consistently chosen by both independent classifiers, lending credibility to the signal.

Equally telling was what did not matter. Wall shear stress, long the centerpiece of hemodynamic risk research, showed no significant differences between groups in univariate comparisons, and only one shear-related feature, the skewness of relative wall shear stress during diastole, was selected by a classifier. The authors argue this makes physical sense: wall shear stress quantification is exquisitely sensitive to imaging resolution, surface reconstruction choices, and near-wall velocity treatment, with reported deviations of up to 38 percent in severely stenotic cases. Bulk intravascular quantities like helicity, by contrast, are more robust to these uncertainties, a lesson already emerging from coronary flow studies.

Geometric features fared no better. Bifurcation angle, tortuosity, and diameter ratios, all previously proposed risk markers, showed no significant group differences and were largely ignored by the classifiers. The team notes that bifurcation angle measurements shift dramatically depending on where along the curving branches they are taken, underscoring the field’s need for standardized geometric definitions, or perhaps a shift away from geometry altogether toward flow-based biomarkers.

The clinical vision is a triage system rather than an MRI replacement. Because neurovascular ultrasound and CT angiography are already part of routine workups, and the simulation and analysis pipeline is largely automated, the workflow could pre-screen patients locally, referring only those flagged as high-risk for specialized plaque MRI while sparing low-risk patients unnecessary trips to tertiary centers. Sub-analyses excluding non-stenotic arteries preserved classifier performance, and the authors emphasize that the ability to correctly classify healthy contralateral vessels is practically important since both sides are routinely assessed.

Caution is warranted. This was a retrospective, single-center proof of concept with only eighteen arteries, all from surgical candidates, and the authors themselves frame the conclusions as hypothesis-generating rather than confirmatory. Prospective, multicentric validation with larger cohorts, automated segmentation, and eventually hard outcomes such as stroke incidence will be essential. Still, the idea that the statistics of swirling blood, computed from scans most clinics already possess, could whisper which plaques are about to turn lethal is a compelling glimpse of accessible, physics-driven stroke prevention.

Subject of Research: Using computational fluid dynamics from CT angiography and ultrasound with machine learning to identify vulnerable carotid plaques

Article Title: Data driven approach to identify vulnerable plaques in carotid atherosclerosis using computed tomography imaging and computational fluid dynamics: A pilot study

Article References: Zheng, S., Bervini, D., Besson, H., Martínez, S. G., Heldner, M. R., Müller, M., Meinel, T., Kaesmacher, J., Harloff, A., & Obrist, D. (2026). Data driven approach to identify vulnerable plaques in carotid atherosclerosis using computed tomography imaging and computational fluid dynamics: A pilot study. PLOS Digital Health, 5(10), e0001762. https://doi.org/10.1371/journal.pdig.0001762

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001762

Keywords: carotid atherosclerosis, vulnerable plaque, computational fluid dynamics, CT angiography, neurovascular ultrasound, machine learning, stroke, plaque MRI, helicity, wall shear stress, hemodynamics, PLOS Digital Health

The research, led by Chuanghui Zhou and Mengting Hu of Southwest Hospital at Army Medical University in Chongqing, China, together with colleagues, set out to answer a deceptively simple question: when you measure a narrowed brain artery with different imaging tools, which tool tells the truth? To find out, the team turned to digital subtraction angiography, or DSA, the long-standing gold standard for assessing blood vessels in the brain. DSA involves threading a catheter into the arterial system, injecting contrast dye directly into the cerebral circulation, and capturing X-ray images in real time. Because it shows contrast flowing through the vessel lumen with exceptional clarity, DSA remains the benchmark against which every non-invasive technique must ultimately be judged, even though it is invasive and carries its own small risks.

The study enrolled 115 patients with non-cardioembolic ischemic stroke, meaning strokes not caused by clots originating in the heart, who were evaluated between September 2022 and September 2025. From this group, the researchers identified 122 stenotic lesions in the middle cerebral artery. Each patient underwent both computed tomography angiography, or CTA, and high-resolution vessel wall imaging on a 7-Tesla MRI scanner. Two blinded radiologists then independently measured the severity of each narrowing, both by diameter and by cross-sectional area, as well as the length of the diseased segment. Blinding is critical in this kind of validation work: if the readers had known what the other modality or the reference standard showed, their measurements could have been unconsciously biased toward agreement.

The technical logic behind the comparison deserves attention. CTA estimates stenosis by imaging contrast-filled blood vessels with X-rays, producing bright, high-contrast pictures of the lumen that can be reconstructed into multiplanar and curved planar views. MRI vessel wall imaging at 7 Tesla works differently. The ultra-high magnetic field delivers a dramatically stronger signal and finer spatial resolution, allowing radiologists to visualize not just the channel through which blood flows but the vessel wall itself, where atherosclerotic plaque accumulates. This matters because the conventional method for grading stenosis, derived from the Warfarin-Aspirin Symptomatic Intracranial Disease trial, compares the narrowest residual lumen diameter with a reference diameter, and small measurement errors at the narrowest point are amplified into large errors in the calculated percentage of stenosis.

When the researchers compared each technique against DSA using the concordance correlation coefficient, a statistic that captures both precision and accuracy, the results favored the ultra-high-field scanner. For diameter-based stenosis measurements, CTA achieved a concordance correlation coefficient of 0.81, while 7-Tesla high-resolution vessel wall imaging reached 0.92. Bland-Altman analysis, which plots the differences between two measurement methods against their average to reveal systematic and proportional bias, told a similar story. CTA showed a mean bias of 0.06 with 95 percent limits of agreement spanning from minus 0.04 to 0.15, whereas the 7-Tesla technique showed a tighter mean bias of 0.02 with limits of agreement from minus 0.07 to 0.10. In plain terms, the MRI measurements clustered more closely around the values that DSA would have produced.

Lesion length, a variable that increasingly influences treatment decisions because long segments of disease may be less suitable for stenting or other endovascular interventions, showed an even more striking gap. The concordance correlation coefficient for length was 0.91 for CTA and 0.94 for vessel wall imaging, but the agreement limits diverged sharply. CTA measurements deviated from DSA by a mean of 0.48 millimeters, with 95 percent limits of agreement stretching from minus 2.35 to 1.40 millimeters. The 7-Tesla MRI measurements showed a mean bias of just 0.03 millimeters, with limits of agreement of minus 1.49 to 1.57 millimeters. For a neurointerventionalist planning a procedure inside an artery only a few millimeters wide, a two-millimeter uncertainty about where the disease begins and ends is not a trivial margin.

The study also explored an alternative way of quantifying narrowing. Instead of the traditional diameter-based calculation, the researchers measured the cross-sectional area of the stenotic lumen on both CTA and MRI and asked how well those area-based values correlated with the diameter-based stenosis percentages derived from DSA. The correlations were strong for both modalities, with Spearman correlation coefficients of 0.79 for CTA and 0.90 for high-resolution vessel wall imaging, both statistically significant. This finding supports the idea that area-based assessment, which some investigators argue better reflects the hemodynamic burden of a stenosis because flow capacity depends on the cross-sectional area of the vessel rather than a single linear diameter, could become a meaningful complement to conventional grading in intracranial disease.

Reproducibility, the question of whether different observers and the same observer on different days will produce the same numbers, is where the new technique truly distinguished itself. Intraclass correlation coefficients for inter-observer and intra-observer agreement ranged from 0.83 to 0.97 for CTA and from 0.94 to 0.99 for 7-Tesla vessel wall imaging. Values approaching 1.0 indicate near-perfect consistency, and the MRI figures sit at the very top of that range. In clinical practice, high reproducibility translates directly into confidence: when two specialists independently grade the same artery and reach the same conclusion, treatment decisions rest on firmer ground, and when a patient is followed over time, true progression of disease can be distinguished from measurement noise.

The implications reach beyond the radiology reading room. Current guidelines for symptomatic intracranial atherosclerotic disease hinge on stenosis severity, with aggressive medical management for most patients and consideration of endovascular treatment in selected severe cases. If the imaging that feeds those decisions underestimates or overestimates the narrowing, patients may be steered toward the wrong therapy. A non-invasive modality that matches the accuracy of catheter angiography could, in principle, spare some patients an invasive diagnostic procedure while giving clinicians richer information about plaque burden and vessel wall pathology that DSA, which shows only the lumen, cannot provide. The 7-Tesla scanner’s ability to image the arterial wall directly opens a window onto the biology of the plaque itself, including features associated with vulnerability and future stroke risk.

Important caveats remain before ultra-high-field MRI can reshape routine stroke care. Seven-Tesla scanners are expensive, less widely available than 1.5- or 3-Tesla systems, and subject to stricter safety screening because of their powerful magnetic fields. The study was retrospective, conducted at a single center with dedicated 7-Tesla infrastructure and specialist support, including collaboration with an industry MR research team, and its findings will need confirmation in broader, multicenter populations. The authors report no competing interests, and the work received support from dedicated 7-Tesla magnetic resonance research funds at Army Medical University. Still, the message of the validation is clear and, for a field that has long accepted the compromises of 3-Tesla imaging, quietly revolutionary: at 7 Tesla, non-invasive vessel wall imaging can measure brain artery narrowing with an accuracy and consistency that rivals, and in several respects exceeds, the CT angiography standard, bringing medicine a step closer to diagnosing the deadliest blockages without a single catheter entering the body.

Subject of Research: Validation of 7-Tesla high-resolution vessel wall MRI against DSA for quantifying middle cerebral artery stenosis

Article Title: Accuracy of 7T high-resolution vessel wall imaging versus CTA in middle cerebral artery stenosis: a DSA-based validation study

Article References: Zhou, C., Hu, M., Xiao, C., Hsu, Y., Chen, W., Chen, P., Liu, H., He, M., & Chen, J. (2026). Accuracy of 7T high-resolution vessel wall imaging versus CTA in middle cerebral artery stenosis: a DSA-based validation study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02864-5

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02864-5

Keywords: 7-Tesla MRI, vessel wall imaging, middle cerebral artery, intracranial stenosis, CT angiography, digital subtraction angiography, ischemic stroke, atherosclerosis, neuroradiology, stenosis grading, medical imaging, reproducibility

The research, led by Valeriia Abramova of the Computer Vision and Robotics Institute at the University of Girona, together with neurologists at Hospital Universitari Dr Josep Trueta, addresses a gap that has limited existing AI stroke tools. Most current detection systems treat the problem as a simple yes-or-no question: is a large vessel occlusion present, or not? But clinicians need more than that. The choice of endovascular strategy differs depending on which vessel segment is occluded. Blockages in the posterior circulation, such as the basilar artery, are more often treated with balloon angioplasty and permanent stents, while occlusions in the anterior circulation typically call for stent thrombectomy or catheter aspiration. Clinical guidelines for posterior circulation occlusions are still evolving, and uncertainty persists over optimal anesthesia and treatment approaches. Knowing the precise vessel involved therefore shapes intervention planning, prognosis estimation, and clinical interpretation in ways a binary alarm cannot.

Commercial software packages such as RapidLVO, Viz-LVO, e-CTA, and StrokeViewer-LVO have brought automated detection into hospitals, but they remain constrained by reduced sensitivity for distal occlusions, vulnerability to imaging artifacts that trigger false positives, and dependence on standardized acquisition protocols that can hamper generalization across institutions. Recent research efforts have pushed toward explicit localization of occlusions, including two-stage convolutional neural network pipelines on four-dimensional CT angiography and self-configuring object detection frameworks applied to maximum intensity projection images. Yet most of these approaches stop at drawing a bounding box around the affected region without linking that region to a specific, named vessel segment. The Girona team’s contribution is to do both tasks at once, within a single unified framework.

Technically, the researchers adapted the state-of-the-art nnDetection framework, a self-configuring three-dimensional object detection method built on the Retina U-Net architecture. This network fuses the RetinaNet one-stage detector with a U-Net encoder-decoder, extracting multi-level image features through convolutional layers with instance normalization and ReLU activation, while lateral and transposed convolutional connections build a feature pyramid. Two sub-networks operate on the pyramid levels: one for classification and one for bounding box regression, trained respectively with binary cross-entropy loss and a generalized intersection-over-union loss, with additional semantic segmentation supervision during training. The classification branch labels each detected occlusion as belonging to one of three clinically crucial segments: the basilar artery, the terminal internal carotid artery, or the M1 segment of the middle cerebral artery. Model training used a fixed patch size of 128 by 128 by 128 voxels, sixty epochs, and stochastic gradient descent with Nesterov momentum, implemented in PyTorch on an NVIDIA GeForce GTX 1080 Ti GPU.

The study’s most distinctive twist lies in how it constrains the search space anatomically. The Circle of Willis is the ring-shaped arterial network at the center of the brain that supplies blood to cerebral tissue, and it encompasses precisely the vessels most often affected by these occlusions. Because manual segmentation of this structure is laborious and ground truth vessel masks were unavailable for the datasets, the team trained a standard nnU-Net segmentation model on data from the TopCoW challenge, a competition dedicated to topology-aware Circle of Willis segmentation in CT and MR angiography spanning diverse anatomical variants. Their segmentation model achieved a Dice score of 0.944 plus or minus 0.026. Applied to the stroke scans, the automatic Circle of Willis segmentation was expanded by 50 voxels in all three dimensions, a margin chosen to guarantee that all occlusions in the dataset fell inside the resulting region of interest.

This anatomical prior enabled a head-to-head comparison of two strategies. In the global approach, the network trained and inferred on the full CT angiography volume. In the local approach, images were cropped to the Circle of Willis-derived region before training and inference. The development dataset comprised 179 CT angiography scans acquired on a Philips Ingenuity scanner at Hospital Dr. Josep Trueta, all containing occlusions annotated with three-dimensional bounding boxes by an expert neurologist. Of these, 143 scans were used for training with five-fold cross-validation, while 36 were held out for internal testing. Class distribution reflected the natural epidemiology of these strokes: the M1 segment dominated at 67 percent of cases, terminal internal carotid artery occlusions comprised 25 percent, and basilar artery occlusions were rare at 8 percent.

The results on the internal test set were strikingly strong for both variants. The global approach achieved a detection sensitivity of 0.92 at 0.08 false positives per image, rising to 0.97 at 0.20 false positives per image, while the local approach matched the same sensitivities at just 0.03 and 0.13 false positives per image respectively. A case-level paired bootstrap analysis at a fixed operating point of 0.1 false positives per image found a mean sensitivity difference of 0.000 with a 95 percent confidence interval spanning negative 0.081 to positive 0.081, confirming no systematic performance gap between the strategies. Each approach missed only a single occlusion. Localization accuracy was equally tight: the mean three-dimensional distance between predicted and ground truth bounding box centers was 1.76 millimeters for both strategies. Interestingly, the full-volume model generated more low-confidence false positives, whereas the region-restricted model produced fewer spurious detections overall.

Classification performance held up nearly as well. On the internal test set, overall accuracy reached 94 percent for the global approach and 91 percent for the local approach, with Cohen’s kappa statistics of 0.88 and 0.82, both indicating almost perfect agreement with expert ground truth. Every basilar artery occlusion was classified correctly by both approaches, a success the authors attribute to the basilar artery’s distinctive, isolated position at the base of the brain, which reduces ambiguity. The main classification shortfall appeared in the terminal internal carotid artery class, where accuracy fell from 89 percent globally to 78 percent locally. Counting detection and classification together, the occlusion was both found and correctly labeled in 33 of 36 cases for the global approach and 32 of 36 for the local one.

Generalization was tested on the independent CODEC-IV benchmark, consisting of 48 CT perfusion-derived CT angiography scans from different scanners and hospitals. Detection sensitivity dropped to 0.71, but much of that decline traced to a mismatch in annotation conventions: the benchmark’s ground truth boxes were uniformly small and pinpointed the occlusion site, while the Girona training boxes captured the entire clot extent. When the true positive threshold was relaxed to 2 percent intersection over union, sensitivity climbed to 0.90 for the global approach and 0.88 for the local one, with mean center-to-center distances of roughly 2.4 millimeters confirming that the models were accurately placing their predictions even when box dimensions diverged. On a supplementary inference-only evaluation of the IACTA-EST challenge dataset, which included 51 occlusion-negative cases, the global model achieved an area under the curve of 0.95 in separating positive from negative cases, against 0.84 for the local model, hinting that a well-chosen confidence threshold could suppress false alarms in clinical triage.

Perhaps the most consequential number is temporal. Restricting inference to the Circle of Willis region made the local approach approximately 3.3 times faster than the global one, cutting mean processing time per case from about 227 seconds to roughly 89 seconds. In acute stroke care, where treatment delays of minutes measurably worsen neurological outcomes, that speedup matters even before accounting for the reduced computational cost of running on modest hospital hardware. The authors acknowledge limitations, including a single-center training cohort from one scanner, a small number of basilar artery cases, and class imbalance mirroring natural disease prevalence. Nonetheless, by coupling a deep object detection framework to an automatically derived vascular landmark, the study demonstrates that anatomical knowledge can be baked directly into machine learning pipelines, delivering detection and vessel-level diagnosis in one pass and pointing toward AI assistants that fit realistically into the breakneck rhythm of a stroke unit.

Subject of Research: Deep learning detection and vessel-level classification of large vessel occlusions in brain CT angiography guided by Circle of Willis localization

Article Title: Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA

Article References: Abramova, V., Oliver, A., Lal-Trehan Estrada, U. M., Hamadache, R. E., Martínez Arias, P., Freixenet, J., Terceño, M., Silva, Y., & Lladó, X. (2026). Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA. Neuroinformatics, 24(4), Article 62. https://doi.org/10.1007/s12021-026-09817-x

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09817-x

Keywords: large vessel occlusion, stroke, CT angiography, Circle of Willis, deep learning, nnDetection, neuroinformatics, medical imaging, endovascular treatment, brain imaging, artificial intelligence, stroke diagnosis