Abhi Jain

Invited talk. Western Neuroradiological Society, Maui, October 1, 2026

From pixels to prognosis

What next-generation AI can already do for carotid plaque, across vessel wall MRI, CTA, and ultrasound.

Download the slides (PDF)Browse the slidesReferences

The argument

We know what matters in a carotid plaque. We just can’t measure it at scale.

Composition predicts stroke better than stenosis does: once lipid-core burden enters a CTA model, stenosis stops being significant. But hand-quantifying the components of a single carotid CTA takes 20 to 30 minutes, and readers agree only moderately, with a kappa of 0.49 on MRI plaque vulnerability.

In 2026, multicenter AI on all three modalities reached the step that matters most: its output now tracks future strokes. The bottleneck was never knowledge. It was measurement.

Three things to remember

  1. Composition is finally measurableIn seconds, reproducibly, on the modalities we already use.
  2. Plaque AI reached outcomesMRI, CTA, and ultrasound models now each link their output to future cerebrovascular events.
  3. The frontier isn’t a better Dice scoreIt’s diverse populations, longitudinal tracking, feeding risk models like IMPROVE, and trials of AI-guided care.

The evidence ladder

Every model has to climb the same rungs: does it measure accurately, does it match biology, does it help a reader, and does it predict what happens to the patient? The last rung, whether AI-guided care changes outcomes, hasn’t been tested yet.

TechnicalDice, ICC, AUC BiologicalHistology, PET, OCT ClinicalReader studies, time PrognosticFuture events InterventionalNot yet MRI CTA Ultrasound

Three modalities, one pattern

Vessel wall MRI

CFAPS. Gai et al., BMC Medicine, 2026. 1,315 patients, 4 centers, 3 vendors, with a prospective PET/MRI arm.

AUC 0.92prospective test set, with plaque segmentation matching human inter-observer agreement
κ 0.73against endarterectomy histology
72 s → 0.5 sper read, and inter-reader agreement rose from 0.49–0.57 to 0.72–0.75
HR 4.3–8.9for ipsilateral stroke over 2 to 4 years

Fully automated from the first step: one network finds the bifurcation, each sequence gets its own encoder, and the plaque mask becomes the prior for the vulnerability call.

The catch

The label is binary, with no component volumes. One country, histology in 42 specimens, and it needs contrast, a dedicated coil, and time. In the stroke bay, we get CTA.

CT angiography

Zhao et al., Academic Radiology, 2026. 2,164 CTAs, 29,113 annotated plaques, 9 centers, and 610 symptomatic patients followed.

76 sper case, fully automated, versus 20 to 30 minutes by hand
ICC 0.86–0.96against radiologist measurements
0.60 → 0.74AUC for recurrent stroke or TIA when AI lipid-core burden is added
HR 1.01for stenosis once lipid core is in the model, no longer significant

The workhorse becomes a composition scanner: vessels are segmented and straightened, plaques detected, and every voxel sorted into lipid core, fibro-lipid, fibrotic tissue, or calcium by density.

The catch

Hemorrhage isn’t reliably separable, because its density overlaps fibrous tissue and contrast. Retrospective, and an AUC of 0.74 can’t decide revascularization. Prevention happens on ultrasound.

Ultrasound

Xu et al., npj Digital Medicine, 2026. 6,618 high-risk participants, 38,090 B-mode images, 4 centers.

0.90AUC, against 0.72 to 0.80 for six experienced sonographers
0.88AUC in external testing
HR 2.09adjusted, for stroke or TIA in the AI high-risk group
277 eventsin 2,174 people over 37 months

An AI that reasons like a sonographer: it must first find plaque and name its features, from echogenicity to surface to stenosis, before it’s allowed to call a plaque vulnerable.

The catch

Incorporation bias, since B-mode reads were part of the reference. Events weren’t adjudicated as ipsilateral, and it’s one country. If it beats us, how do we know why?

How these models earn trust

Saliency: where it lookedMRI heat maps land on hemorrhage, ulceration, and lipid core, and stay diffuse on stable plaque. But a heat map isn’t a reason.
Glass box: how it reasonedThe ultrasound model has to name features first. Nearly every vulnerable call carried irregular morphology, yet the full model beat any single feature.
Measurement: what it measuredCTA outputs cubic millimeters of lipid core, not just a probability. You can inspect it, quantify it, and audit it.
Biology: whether it’s realPlaques the MRI model calls vulnerable match histology and burn hotter on FDG-PET.

Shared blind spots

One countryAll three anchor cohorts come from China. Diverse, Western validation is still needed.
We’re still the truthLabels are mostly radiologist reads, with histology in 42 specimens and OCT in 16 patients.
Association isn’t interventionNo one has yet randomized AI-guided care against standard management.
Endpoints don’t matchIpsilateral stroke, recurrent stroke or TIA, and overall stroke or TIA, with no head-to-head comparison.

Notice the teacher: vessel wall MRI validated the CTA model, and contrast ultrasound, CTA, and MRI helped label the ultrasound data. Advanced imaging is becoming the teacher for routine imaging.

Where it’s heading

From snapshot to trajectoryDeep learning on serial vessel wall scans over about six years: plaques without hemorrhage barely changed, while those with it grew, and new hemorrhage persisted.
AI before the readDenoising and super-resolution nearly doubled in-plane resolution on intracranial vessel wall MRI, and hemorrhage sensitivity in culprit plaques rose from 24% to 71%.
Open weightsThe CFAPS weights and the ultrasound and vessel wall code are public, so we can validate them on our own patients.
The missing rungIMPROVE turns MRI hemorrhage into a three-year stroke risk. AI can standardize that input. Then we randomize.

The slides

References

  1. Gai Q, Zhang Y, Liu J, et al. A fully automated deep learning pipeline system for carotid plaque segmentation and vulnerability classification on multi-contrast MRI: a multicenter study. BMC Med. 2026. doi:10.1186/s12916-026-05199-8
  2. Zhao MJ, Cheng XQ, Xu PP, et al. Automated analysis of carotid plaque components in CT angiography: performance and prognosis in symptomatic carotid stenosis patients. Acad Radiol. 2026. doi:10.1016/j.acra.2026.06.066
  3. Xu Q, Ding H, Zhou L, et al. AI-derived carotid plaque vulnerability from B-mode ultrasound for cerebrovascular risk stratification: a multicenter cohort study. npj Digit Med. 2026. doi:10.1038/s41746-026-03201-4
  4. Guo Y, Akcicek EY, Hippe DS, et al. Long-term carotid plaque progression and the role of intraplaque hemorrhage: a deep learning-based analysis of longitudinal vessel wall imaging. J Cardiovasc Magn Reson. 2026;28:102670. doi:10.1016/j.jocmr.2025.102670
  5. Fu F, Lin Z, Yang X, et al. Deep learning for high-resolution MR vessel wall imaging: image reconstruction, stenosis diagnosis and plaque calculation. Eur Radiol. 2026;36:5169–5181.
  6. Yang L, Yang X, Gong Z, et al. Deep learning-based automatic segmentation of arterial vessel walls and plaques in MR vessel wall images for quantitative assessment. Eur Radiol. 2025;35:7712–7726.
  7. Seo M, Jung W, Jeong G, et al. Deep learning improves quality of intracranial vessel wall MRI for better characterization of potentially culprit plaques. Sci Rep. 2024;14:18983.
  8. Nies KPH, Smits LJM, van Kuijk SMJ, et al. Individualized MRI-based stroke prediction score using plaque vulnerability for symptomatic carotid artery disease patients (IMPROVE). Stroke. 2025;56:2068–2078. doi:10.1161/STROKEAHA.124.050020
  9. Ma D, Chen J, Li Z, et al. Carotid plaque vulnerability assessment and stroke risk prediction based on radiomics and deep learning: a systematic review and meta-analysis. BMC Med Imaging. 2026. doi:10.1186/s12880-026-02824-z
  10. Ju L, Guo Y, Guo H, et al. Diagnostic performance of deep learning and radiomics in extracranial carotid plaque detection: systematic review and meta-analysis. J Med Internet Res. 2026;28:e77092. doi:10.2196/77092