techbullion.com

Inside AtheroPoint’s Artificial Intelligence Push to Map Plaque Burden in Neck Arteries

By Victor Miller·Published November 18, 2026·HealthTech, Technology

AtheroPoint researcher

In an October 2026 benchmark, AtheroPoint’s AtheroEdge 2.0 analysed a neck ultrasound image in six milliseconds on an NVIDIA L40 GPU. Across 400 images from 200 patients in Japan, it recorded a wall-thickness error of 0.011 millimetres against expert tracings, the lowest of the four systems tested. The other three used deep learning.

AtheroPoint, a medical imaging company based in Roseville, California, builds software that turns ultrasound images into plaque measurements. Its desktop-based product, AtheroEdge, automatically traces the lumen-intima and media-adventitia (LIMA) borders and measures the plaque burden in these artery walls. AtheroCloud brings automated measurements to a web browser. In parallel, the company’s Artificial Intelligence (AI) research explores how to outline plaque borders and calculate the plaque burden it covers in these ultrasound scans.

How AtheroEdge finds the two faint lines

The carotid arteries carry blood to the brain. On a neck ultrasound, the channel where blood flows often looks dark, while the surrounding tissue appears in shades of grey, which includes intima layers and media layers. AtheroEdge traces two boundaries (called LIMA borders) on the artery wall farthest from the ultrasound probe. The distance between those lines is called carotid intima-media thickness, or cIMT. The software measures it at multiple points and reports how it varies along the walls of the artery.

Finding those lines is a computer vision problem. On a grainy scan, a change in brightness might be the edge of a tissue layer or simply image noise or sometimes black blob called shadow. An early version of AtheroEdge smoothed out abrupt jumps (spike) in the outline. Researchers tested it on 885 low-contrast images from Asian’s continent. The software’s average reading was 0.407 millimetres, compared with 0.429 millimetres from a vascular radiologist’s tracings.

The original AtheroEdge received U.S. FDA 510(k) clearance in 2012 for automated measurement of carotid intima-media thickness from ultrasound images. The software accepts standard DICOM medical image files and generates a report physicians can review alongside the patient’s other medical information, which is stored in the database management system (DBMS). The reports can be emailed to the doctor or patient directly. Further AtheroEdge runs on the routine mode (while the patient is with the doctor) or pharmaceutical mode (where the system is run over 5000 ultrasound scans in just 15,000 seconds, given in the whole set of wall parameters like cIMT max, min, average, IMT variability (IMTV), total plaque area (TPA), and maximum plaque height (MPH)).

A 2017 study tackled where along the artery to take that measurement. The software found the carotid bulb, where the artery widens before branching, and measured the wall in sections at set distances from that landmark. The team tested the approach on 649 ultrasound scans from different imaging scanners from different original equipment manufactures (OEM) vendors.

What AtheroPoint’s Artificial Intelligence system learns from expert tracings

Plaque can form an irregular patch in the artery wall. To measure the area it covers, researchers need an outline that follows its topology and shape. In a 2022 study that included 970 images from the UK, a team co-led by AtheroPoint’s Jasjit S. Suri trained a neural network on scans marked by experts. The network learned to identify which pixels belonged inside the LIMA outline. A technique called attention gave more weight to image features that helped it find the boundaries.

In August 2026, a study in Diagnostics tested 12 versions of a two-pass (cascaded systems) approach. The first pass sketched the artery wall. The second used both that sketch and the original scan to refine the outline by following the plaque topology more accurately. The researchers trained the models on scans from Japan, then tested them on a separate Hong Kong dataset without retraining. They also checked the outlines against expert tracings on example scans from two European centres.

The October benchmark showed what a two-stage design can deliver and what it costs to train. Among the three AI models, the two-stage model with attention came closest to the experts’ outlines. It achieved a Dice score of 91.2%, which measures how closely the areas marked by the AI model and the experts overlap. Its wall-thickness error was 0.022 millimetres, and it used about 1.7 times as much graphics memory during training as the model without attention. AtheroEdge 3.0 used edge detection and produced the closest thickness readings in the benchmark.

Thickness and overlap answer different questions. If both traced borders shift together, the distance between them can still be correct. Checking the overlap shows whether the software traced the same patch of tissue, which matters when researchers use those outlines to estimate the plaque area.

AtheroCloud takes the analysis online

With AtheroCloud, a clinic can upload ultrasound images to AtheroPoint’s cloud for processing. In a published pilot, clinicians ran automated wall-thickness measurements through a browser and saved the reports online. Researchers tested the workflow on 200 scans from 100 patients using Toho University Japan data sets. On both sides of the neck, the software’s average readings were within 0.03 millimetres of the averages from a radiologist’s manual tracings.

A later pilot used AtheroCloud to measure the width of the channel where blood flows. Operators repeated the analysis to see how consistent the readings were. For a trial spanning multiple hospitals, that workflow gives researchers a way to collect scans locally and apply the same measurement method across sites. In a related AtheroCloud study, operators used scale markers on the scans to convert pixel distances into millimetres so the measurements could be compared.

The October benchmark, which Suri co-authored, drew on a single Japanese patient group. The authors called for testing at more hospitals and in more varied populations. For studies that track patients over time, one practical test is whether two scanners give comparable readings of the same artery. That consistency helps researchers judge whether a small difference between visits reflects a change in the artery or a change in how it was scanned. AtheroPoint’s AtheroEdge was clinically validated by comparing Ultrasound-based measurements against Computed Tomography (CT)-based measurements, showing comparable results.

References

2026 AtheroEdge 2.0 and deep learning benchmark (Cardiovascular Diagnosis and Therapy) AtheroEdge product overview (AtheroPoint)

AtheroEdge 510(k) clearance record (K122022) (U.S. Food and Drug Administration) AtheroEdge 510(k) technical summary (U.S. Food and Drug Administration)

2012 AtheroEdge study of 885 low-contrast ultrasound images (International Angiology)

2017 study of carotid wall measurements across scanners (Computer Methods and Programs in Biomedicine) 2022 study of attention-based AI for plaque segmentation (Journal of Cardiovascular Development and Disease) 2026 study of two-stage AI for carotid wall segmentation (Diagnostics)

2016 AtheroCloud pilot on carotid wall thickness (Computers in Biology and Medicine) 2018 AtheroCloud study of carotid lumen diameter (Indian Heart Journal)

2018 AtheroCloud study of measurement reproducibility (Journal of Clinical and Diagnostic Research)