The Sound of Death: Deep Learning Reveals Vascular Damage from Carotid Ultrasound

Christoph Balada; Aida Romano-Martinez; Payal Varshney; Vincent ten Cate; Katharina Geschke; Jonas Tesarz; Paul Claßen; Alexander K. Schuster

cs.LGFeb 19, 2026

The Sound of Death: Deep Learning Reveals Vascular Damage from Carotid Ultrasound

Authors: Christoph Balada, Aida Romano-Martinez, Payal Varshney, Vincent ten Cate, Katharina Geschke, Jonas Tesarz, Paul Claßen, Alexander K. Schuster, +5 more

Organizations: German Research Center for Artificial Intelligence (DFKI), 67663 Kaiserslautern, Germany · RPTU University Kaiserslautern-Landau, 67663 Kaiserslautern, Germany · University Medical Center of the Johannes Gutenberg-University Mainz, Germany · German Center for Cardiovascular Research (DZHK), Germany

Abstract

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, yet early risk detection is often limited by available diagnostics. Carotid ultrasound, a non-invasive and widely accessible modality, encodes rich structural and hemodynamic information that is largely untapped. Here, we present a machine learning (ML) framework that extracts clinically meaningful representations of vascular damage (VD) from carotid ultrasound videos, using hypertension as a weak proxy label. The model learns robust features that are biologically plausible, interpretable, and strongly associated with established cardiovascular risk factors, comorbidities, and laboratory measures. High VD stratifies individuals for myocardial infarction, cardiac death, and all-cause mortality, matching or outperforming conventional risk models such as SCORE2. Explainable AI analyses reveal that the model relies on vessel morphology and perivascular tissue characteristics, uncovering novel functional and anatomical signatures of vascular damage. This work demonstrates that routine carotid ultrasound contains far more prognostic information than previously recognized. Our approach provides a scalable, non-invasive, and cost-effective tool for population-wide cardiovascular risk assessment, enabling earlier and more personalized prevention strategies without reliance on laboratory tests or complex clinical inputs.

Figures & tables

Figure 1: Classification accuracy by age at initial assessment. Left: male participants; right: female participants. Dark colours denote correct classifications, with true positives (high vascular damage (VD) and hypertensive) shown in red and true negatives (low VD and non-hypertensive) in blue. Light shades indicate misclassifications, specifically false positives (light red; high VD and non-hypertensive) and false negatives (light blue; low VD and non-hypertensive). Overall, classification patterns between males and females are comparable above 55 years of age. However, among younger individuals (<45 years), males exhibit a higher proportion of false negatives, accompanied by a reduced rate of true negatives compared with females.

Figure 2: Average, median, top-5% positive and negative regions of interest (ROIs) of 50 videos. ROIs highlight which regions have been found most important for the model to predict an example as high vascular damage (green) or as low vascular damage (red). Attributions and background images have been averaged over 50 videos, selected around the global dataset mean. Progress in course of the different assessments is illustrated per row. In all cases the vessel walls are rendered most important. However, in particular in older ages (e.g. 10-year follow-up), ROIs are in particular pronounced in the perivascular tissue.

Figure 3: Comparison of clinical parameters in the Gutenberg Health Study cohort. Higher values generally indicate poorer cardiovascular status. Individuals with high vascular damage, irrespective of hypertension, consistently exhibit the poorest overall condition.

Panel of Figure 4 — Figure 4: Kaplan–Meier estimates for vascular damage (VD). Top: (Event-free) survival for incident myocardial infarction, stroke, and cardiac death over a 5-year period (left) and for cardiac and all-cause death over 10 years (right), comparing participants with high VD (solid lines) versus low VD (dashed lines). Bottom: 15-year all-cause death stratified by VD and hypertension label used during model training (left) and by VD versus SCORE2 baseline risk model (right). VD consistently outperformed hypertension status. High VD at the predefined threshold of $0.67$ exhibited risk-stratification characteristics, closely comparable to a high SCORE2.

Panel of Figure 5 — Figure 5: Cox proportional hazards models incorporating different covariates. (a) Hazard ratios (left) and corresponding concordance and significance (right) The shaded regions highlight the difference between the lowest and highest C-index. Combining high and low vascular damage score and SCORE2, shows the best performance. (b) Differences in concordance index (C-index) (left) and statistical significance (right).

Figure 6: Left: Receiver operating characteristic (ROC) curves for a linear classifier predicting 5-year cardiac death based on the predicted hazard from the vascular damage (VD) and SCORE2 Cox models. Right: Integrated Discrimination Improvement (IDI) and Net Reclassification Improvement (NRI) comparing both models. The VD-based model demonstrated superior discrimination, achieving higher area under the ROC curve, as well as improved IDI and NRI, indicating enhanced reclassification performance for the fatal 5-year cardiac death subgroup.

Model	$bACC_{val}\uparrow$	$F_{1,val}\uparrow$	$AUC_{ROC,val}\uparrow$
VideoMAE	70.61%	0.733	0.773
ViViT	70.88 %	0.709	0.772
TimeSformer	69.07%	0.724	0.763

Table 2: Individual-level accuracy and balanced accuracy on the validation set for different backbone architectures

Explore similar work

A CUBS-Compatible Ultrasound Morphology and Uncertainty-Aware Baseline for Carotid Intima-Media Segmentation and Preliminary Risk Prediction
May 14, 2026Aueaphum AueawatthanaphisutUltrasound Image Segmentation Clinical Risk Prediction
School of Information, Computer, and Communication Technology Sirindhorn International Institute of Technology, Thammasat University Pathum Thani, Thailand
Mammography Foundation Models for Opportunistic Prediction of Major Adverse Cardiovascular Events
Sep 16, 2026Paula Feldman, Nusrat Binta Nizam, Sunwoo Kwak +3Medical Imaging Foundation Models Clinical Outcome Prediction
Department of Radiology, Weill Cornell Medicine, New York, NY, USA · Cornell Tech, New York, NY, USA · Cornell University, New York, NY, USA
Ultrasound-Based Prediction of Cirrhosis Decompensation Using Large-Scale Computer Vision Models
Sep 3, 2026Guangyi Zhang, Peiyun Ni, Eugene Cheah +4Clinical Outcome Prediction Clinical Risk Prediction
Center for Ultrasound Research & Translation, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA · Division of Gastroenterology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA · Department of Medicine, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA

May 14, 2026J/K move · Enter open · S save

A CUBS-Compatible Ultrasound Morphology and Uncertainty-Aware Baseline for Carotid Intima-Media Segmentation and Preliminary Risk Prediction

Aueaphum Aueawatthanaphisut

School of Information, Computer, and Communication Technology Sirindhorn International Institute of Technology, Thammasat University Pathum Thani, Thailand

Ultrasound Image Segmentation Clinical Risk Prediction

Open paper AskPDF arXiv

Carotid atherosclerosis is a major contributor to ischemic stroke and transient ischemic attack. Conventional ultrasound assessment is commonly based on intima-media thickness, plaque appearance, stenosis degree, and peak systolic velocity, but these morphology- and velocity-based indicators may not fully capture patient-specific vascular risk. This study presents AtheroFlow-XNet, a CUBS-compatible ultrasound morphology and uncertainty-aware learning baseline for carotid intima-media segmentation and preliminary risk prediction. Using the Carotid Ultrasound Boundary Study dataset, manual lumen-intima and media-adventitia boundary annotations were converted into dense intima-media masks for supervised segmentation. Clinical variables were incorporated into an auxiliary risk-prediction branch, and Monte Carlo dropout was used for uncertainty-aware inference. The model was evaluated using a patient-level train-validation-test split with 1,522 training images, 326 validation images, and 328 testing images. The proposed model achieved a Dice coefficient of 0.7930 for LI-MA mask segmentation, a segmentation loss of 0.2359, and an area under the receiver operating characteristic curve of 0.6910 for preliminary risk prediction. Qualitative results showed that predicted masks were generally aligned with manual annotations, while uncertainty maps highlighted ambiguous wall-boundary regions. These results suggest that ultrasound-derived carotid morphology can support automated wall analysis and uncertainty-aware interpretation. Since CUBS does not provide Doppler waveforms or CFD-derived hemodynamic biomarkers, this work should be interpreted as a reproducible morphology-driven baseline. Future work will incorporate Doppler-derived flow profiles, patient-specific vascular reconstruction, and CFD-based wall shear biomarkers.

Abstract

Explore similar work

A CUBS-Compatible Ultrasound Morphology and Uncertainty-Aware Baseline for Carotid Intima-Media Segmentation and Preliminary Risk Prediction

Mammography Foundation Models for Opportunistic Prediction of Major Adverse Cardiovascular Events

Ultrasound-Based Prediction of Cirrhosis Decompensation Using Large-Scale Computer Vision Models