AI-Enabled Carotid Stenosis Assessment

This paper presents a method for evaluating ischemic stroke risk through the quantification of carotid artery stenosis using deep learning techniques. By employing U-Net architecture, we achieve accurate segmentation of carotid arteries from longitudinal and transverse 2D ultrasound images. This allows for precise measurement of artery diameter and lumen area reduction. Furthermore, we produce a comprehensive 3D model of carotid artery, enhancing the assessment of stenosis severity, a major contributing factor to ischemic strokes, which constituted 65.3% of all new strokes globally in 2025. Our AI-powered approach enhances consistency and accuracy, offering significant improvements over traditional manual methods, potentially aiding in timely intervention and stroke prevention.

Introduction

Medical imaging has evolved from basic 2D images to advanced AI systems capable of detecting and quantifying diseases with high precision. A key application is the segmentation of carotid arteries in ultrasound images to assess stroke risk. This paper examines how the U-Net architecture enables precise artery segmentation and quantitative analysis of vascular health by computing the percentage reduction in artery diameter and lumen area, vital for evaluating carotid artery stenosis.

Radiologists utilize carotid ultrasound in both transverse and longitudinal views to evaluate ischemic stroke. The transverse view helps identify and describe the plaque, while the longitudinal view measures its impact on blood flow. Together, these complementary views provide the necessary insights for a comprehensive stroke risk assessment.

Clinical Relevance of Carotid Stenosis Quantification

Ischemic stroke constituted 65.3% of all new strokes globally in 2025 around the world. These strokes occur when a blockage in a blood vessel restricts oxygen-rich blood flow to the brain. One of the major contributors to this condition is carotid artery stenosis—caused by plaque buildup that narrows the arteries supplying blood to the brain.

Quantifying this narrowing through metrics such as the percentage reduction in artery diameter and lumen area is critical for evaluating stenosis severity. Accurate, objective measurements support timely intervention, potentially preventing strokes and minimizing long-term disability.

Methodology

The Traditional Workflow: Manual & Subjective

Traditionally, carotid stenosis is assessed by radiologists interpreting ultrasound images, which is manual and subjective. While effective, this process has limitations:

Time-consuming
Subject to human error
Dependent on operator experience

Deep learning addresses these challenges by offering an automated solution that delivers fast, accurate, and consistent results using semantic segmentation.

a) Dataset and Preprocessing:

The dataset contains ultrasound images of the common carotid artery. Original recordings were saved as DICOM time series and later converted into PNG format. These images were carefully cropped to highlight the area of interest. Each subject contributed 100 images, resulting in a dataset of 1,100 images, all annotated with expert labeled masks.

Standardized Preprocessing Pipeline:

Resizing all images to 256×256 pixels.
Normalization of pixel intensities to a 0–1 range.
Grayscale mask conversion, ensuring clarity in semantic classes (vessel vs. lumen).

b) Model

Key Architectural Features of U-Net:

The U-Net model is built for binary image segmentation using a symmetric encoder-decoder architecture with convolutional layers. The encoder comprises four blocks with two Conv2D layers (ReLU, same padding) and MaxPooling2D, progressively increasing filters. A bottleneck with two Conv2D layers connects to the decoder, which up samples using UpSampling2D and skip connections, reducing filters. A final Conv2D (1, 1) layer with sigmoid activation outputs the binary segmentation mask.

The U-Net model is used for segmentation of Carotid artery at various cross sections to get sectional information like accurate vessel and lumen segmentation.

c) Quantification

The true power of semantic segmentation lies not just in outlining anatomy, but in enabling objective quantitative analysis. Once the vessel wall and lumen are segmented, clinically significant vascular metrics can be extracted.

i) Percentage Diameter Reduction

This metric compares the diameter at the most narrowed point with a proximal, healthy segment providing an indicator of stenosis severity.

ii) Percentage Area Reduction

This compares the cross-sectional area of the entire vessel with that of the lumen, giving a more comprehensive view of occlusion, especially when diameter may be distorted by image artifacts or imaging angles.

With accurate segmentation of the vessel and lumen, clinicians can automatically:

Grade stenosis severity using standardized thresholds.
Track disease progression over time.
Decide on interventions (e.g., stenting or surgery).
Predict stroke risk with greater objectivity and consistency.

This automated analysis complements traditional diagnostic methods, reducing subjectivity and enhancing decision-making, especially in high-volume clinical settings.

d) Transverse View Slices

The approach to collect the slice-by-slice segments of the transverse view of Common Carotid Artery (CCA) is represented below:

Conclusion

Cyient proposed U-Net-based segmentation pipeline helps rebuild the 3D model of the Common Carotid artery with exact location of the plague and vessel walls and facilitates automated diagnostics. With quick analysis of vascular structures and real-time computation of stenosis metrics, AI can aid in transitioning stroke care from reactive to preventive. Such models can be integrated into portable ultrasound devices, used in low-resource settings, and assist both specialists and general practitioners in making timely and accurate decisions.

References

Momot, Agata (2022), “Common Carotid Artery Ultrasound Images”, Mendeley Data, V1, doi: 10.17632/d4xt63mgjm.1
Biswas, M., Saba, L., Omerzu, T., Johri, A. M., Khanna, N. N., Višković, K., … & Suri, J. S. (2021). A review on joint carotid intima-media thickness and plaque area measurement in ultrasound for cardiovascular/stroke risk monitoring: artificial intelligence framework. Journal of Digital Imaging. https://doi.org/10.1007/s10278-021-00461-2
Najmath Ottakath, Younes Akbari, Somaya Ali Al-Maadeed, Ahmed Bouridane, Susu M. Zughaier, Muhammad E.H. Chowdhury, Bi-attention DoubleUNet: A deep learning approach for carotid artery segmentation in transverse view images for non-invasive stenosis diagnosis, Biomedical Signal Processing and Control, Volume 94, 2024, 106350, ISSN 1746-8094, https://doi.org/10.1016/j.bspc.2024.106350
Inamdar, V., & Pranathi, G. (2023), “Segmentation of the Carotid Artery Using Deep Learning U-Net Technique.” https://api.semanticscholar.org/CorpusID:269138969
Zhou, J., Zhang, Y., et al. Intelligent Segmentation of Intima–Media and Plaque Recognition in Carotid Artery Ultrasound Images. Published in Medical Physics, 2021. Available at: https://pubmed.ncbi.nlm.nih.gov/34872788
Mao, C.-M., Pang, J.-S., & Lo, Y.-T. (2000) Segmentation of Carotid Artery in Ultrasound Images: Method Development and Evaluation Technique DOI: 10.1118/1.1287111 · Source: PubMed.
Jain PK, Dubey A, Saba L, Khanna NN, Laird JR, Nicolaides A, Fouda MM, Suri JS, Sharma N. Attention-Based UNet Deep Learning Model for Plaque Segmentation in Carotid Ultrasound for Stroke Risk Stratification: An Artificial Intelligence Paradigm. Journal of Cardiovascular Development and Disease. 2022; 9(10):326. https://doi.org/10.3390/jcdd9100326
Kristen M. Meiburger, Francesco Marzola, Guillaume Zahnd, Francesco Faita, Christos P. Loizou, Nolann Lainé, Catarina Carvalho, David A. Steinman, Lorenzo Gibello, Rosa Maria Bruno, Ricarda Clarenbach, Martina Francesconi, Andrew N. Nicolaides, Hervé Liebgott, Aurélio Campilho, Reza Ghotbi, Efthyvoulos Kyriacou, Nassir Navab, Maura Griffin, Andrie G. Panayiotou, Rachele Gherardini, Gianfranco Varetto, Elisabetta Bianchini, Constantinos S. Pattichis, Lorenzo Ghiadoni, José Rouco, Maciej Orkisz, Filippo Molinari, Carotid Ultrasound Boundary Study (CUBS): Technical considerations on an open multi-center analysis of computerized measurement systems for intima-media thickness measurement on common carotid artery longitudinal B-mode ultrasound scans,,2022, https://doi.org/10.1016/j.compbiomed.2022.105333
World Stroke Organization: Global Stroke Fact Sheet 2025 https://pmc.ncbi.nlm.nih.gov/articles/PMC11786524/
Molinari F, Meiburger KM, Zeng G, Saba L, Rajendra Acharya U, Famiglietti L, Georgiou N, Nicolaides A, Sriswan Mamidi R, Kuper H, Suri JS. Automated carotid IMT measurement and its validation in low contrast ultrasound database of 885 patient Indian population epidemiological study: results of AtheroEdge™ Software. Int Angiol. 2012 Feb;31(1):42-53. PMID: 22330624; PMCID: PMC3504971.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Vol 9351. Springer; 2015:234-241. doi:10.1007/978-3-319-24574-4_28
Jiang H, Spence JD, Chiu B. Segmentation of carotid vessel wall using U-Net and segmentation average network. arXiv. Published February 26, 2020. Accessed July 1, 2025. https://arxiv.org/abs/2002.11467
Lainé AF, Cheng K, Moura L, et al. Carotid artery wall segmentation in ultrasound image sequences using a deep convolutional neural network. arXiv. Published January 27, 2022. Accessed July 1, 2025. https://arxiv.org/abs/2201.12152
Deng L, Tian X, Jiang D, et al. Automatic segmentation of ultrasound images of carotid atherosclerotic plaque based on Dense-UNet. Technol Health Care. 2023;31(1):1-11. doi:10.3233/THC-220152
Yu L, Zhou Y, Li Z, et al. Deep learning-based automatic segmentation and evaluation of carotid atherosclerotic plaque in ultrasound images: A review. Neurocomputing. 2023;524:126298. doi:10.1016/j.neucom.2023.126298.
Meiburger, Kristen M.; Zahnd, Guillaume; Faita, Francesco; Loizou, Christos; Carvalho, Catarina; Steinman, David; Gibello, Lorenzo; Bruno, Rosa Maria; Marzola, Francesco; Clarenbach, Ricarda; Francesconi, Martina; Nicolaides, Andrew; Campilho, Aurelio; Ghotbi, Reza; Kyriacou, Efthyvoulos ; Navab, Nassir; Griffin, Maura; Panayiotou, Andrie; Gherardini, Rachele; Varetto, Gianfranco; Bianchini, Elisabetta; Pattichis, Constantinos ; Ghiadoni, Lorenzo; Rouco, José; Molinari, Filippo (2021), “DATASET for "Carotid Ultrasound Boundary Study (CUBS): an open multi-center analysis of computerized intima-media thickness measurement systems and their clinical impact"”, Mendeley Data, V1, doi: 10.17632/fpv535fss7.

About the Authors

Srinivas Rao Kudavelly
Consultant Senior Principal - Healthcare and Life Sciences
Srinivasrao.Kudavelly@cyient.com

Srinivas has over 25 years experience spanning Consumer Electronics, Biomedical Instrumentation and Medical Imaging . He has led research and development teams, focussed on end-to-end 3D/4D quantification applications and released several "concept to research to market" solutions. He led a cross functional team to drive applied research , product development , human factors team, clinical research , external collaboration and innovation. He has garnered diverse sets of skill sets and problem challenges. and has over 30 Patent filings and 15 Patent Grants across varied domains , mentored over 30+ student projects , been a guide for over 10+ master thesis students ,peer reviewer for papers and an IEEE Senior Member ( 2007)

Venkat Sudheer Naraharisetty
Lead Data Scientist - Healthcare and Life Sciences
Venkatsudheer.Naraharisetty@cyient.com

With a robust career spanning over 15 years, he has amassed extensive experience in diverse fields such as Automotive Research and Development, Computer Aided Engineering, and Data Science, with a particular focus on Artificial Intelligence. His expertise extends across a wide array of domains, including crash analysis and the development of classification models using advanced machine learning and deep learning techniques. These skills have been applied in various sectors, notably Automotive, consumer electronics and Medical Imaging.

Possible Challenges that Deployers/Manufacturers/Providers of AI for Medical Purpose Might Face

Integrating AI into medical devices presents complex compliance challenges, particularly due to the dual regulatory landscape governed by the EU AI Act-2024/1689, and the EU MDR-2017/745 or EU IVDR-2017/746. The intersection of AI functionality and medical safety introduces both technical and procedural hurdles.

Dual Regulatory Burden

Manufacturers must comply with both the EU MDR/IVDR and EU AI Act.
Requires harmonization of conformity assessments technical documentation and quality system across two frameworks.
Potential for duplicated or conflicting requirements.

High Risk AI classification

AIMD classified as High Risk as per EU AI Act shall comply stringent requirements such as human oversight, transparency, robustness and post-market monitoring

Data Governance & Quality

Ensuring training and test data are: Relevant, representative, free of bias, Statistically appropriate, Collected lawfully under GDPR
Challenge: Medical AI often uses retrospective, non-standardized, or anonymized datasets with quality or bias issues.

Transparency & Explainability

AI outputs must be understandable to the intended user (e.g., physicians).
Difficult for complex models like deep learning or black-box AI to meet these criteria.
Manufacturers may need to redesign interfaces or limit the use of opaque algorithms.

Human Oversight

Manufacturers must ensure that human operators can understand, intervene, or override the AI system.
Manufacturer must have explicit design features, documentation, and training materials to enable human oversight.

Robustness, Accuracy & Cybersecurity

AI systems must maintain accuracy and performance throughout their lifecycle.
Real-world data often differs from training environments, impacting performance.
Cybersecurity risks increase with connected AI-enabled medical devices (e.g., remote updates, API access).

Continuous Learning & Software Updates

Many AI systems are adaptive or continuously learning, which conflicts with the static certification model under EU MDR and the EU AI Act.
Challenge: Validation & monitoring of that changes postdeployment.

Technical Documentation & Traceability

Manufacturers must produce detailed documentation about:
Model architecture
Training/testing data
Risk management
Logging and traceability mechanisms
These are not always readily available for third-party or opensource AI components.

Post-Market Monitoring

Requires active surveillance of AI model performance, bias, and safety over time.
Need for data pipelines, feedback loops, and incident reporting mechanisms tailored to AI.

Notified Body Expertise Gaps

Many Notified Bodies lack AI-specific expertise.
Manufacturers may face delays or inconsistencies in conformity assessment procedures.
Ongoing need for capacity building among conformity assessment entities.

Ethical and Fundamental Rights Compliance

AI must respect human dignity, privacy, non-discrimination, and autonomy (restricted from AI prohibited practices as per EU AI Act). In case of High-Risk AI model/ system it becomes complicated to demonstrate the above features for the AI involved in life-and-death decisions, triage, or behavior prediction.

Cost & Time of Compliance

Compliance adds substantial regulatory, engineering, and legal costs.
Small or medium manufacturers may find this particularly resource-intensive.

Conclusion

The EU AI Act (2024/1689) marks a transformative step in establishing a robust and harmonized robust regulatory framework for artificial intelligence across the European Union. For the healthcare sector particularly medical device manufacturers integrating AI, this regulation introduces not only new compliance obligations but also opportunities to drive innovation within a well-defined legal and ethical structure.

By adopting a risk-based approach, the Act ensures that AI systems, especially those used in critical sectors like healthcare, are subject to appropriate oversight and accountability. It mandates transparent, safe, and human-centric AI while promoting public trust and technological progress.

To meet these evolving expectations, manufacturers, developers, and deployers of AI systems must align their internal processes with both existing medical device regulations (e.g., EU MDR/IVDR) and AI-specific obligations under this new law.

Proactive compliance will involve:

Integrating regulatory requirements early in the design and development phase
Investing in technical documentation and risk governance
Leveraging recognized standards to establish traceability and conformity
Promoting AI literacy and human oversight across all operational levels

Ultimately, the EU AI Act not only safeguards individuals but also lays the foundation for sustainable and responsible digital health innovation, supporting the ethical use of AI while enabling Europe to lead in the global AI landscape.

About the Authors

Sathish Kumar Thiagarajan is a seasoned Controls & Automation Engineer with over 18 years of global experience in managing large-scale industrial automation projects involving PLCs, SCADA, and Drives. He specializes in optimizing technical workflows, ensuring regulatory compliance, and leading cross-functional teams to deliver seamless IT/OT integration solutions. Known for enhancing operational efficiency and driving cost-effective innovations, his expertise helps shape transformative strategies in industrial automation.

Srinivasu Parupalli is an experienced Systems Engineer with expertise in program management and delivery across multiple domains, including Industry 4.0, Manufacturing, Embedded Systems, IoT, Software Applications Development, and Cloud Integrations. He has extensive experience in end-to-end product development and has been instrumental in building and training teams on emerging technologies such as Ignition, Solumina, Aveva, and SCADA systems for deployment in diverse customer projects. With a strong background in industrial automation, he has worked across various industries, including Manufacturing, Energy, Utilities, Healthcare, and Process Automation, developing MES, SCADA, and HMI solutions integrated with other applications. His expertise lies in customer engagement, requirements analysis, and risk management, ensuring the successful execution of complex automation projects.

About the Author

Abhishek Kumar
Subject Matter Expert in Medical Device Regulatory and Quality Assurance

Abhishek Kumar is a Subject Matter Expert in Medical Device Regulatory and Quality Assurance with over 14 years of experience. He has led the EU MDR 2017/745 sustenance program, managed multiple global engagements for top medical device companies, and supported the gap assessment, remediation, and submission of 70+ technical documents across EU MDR, ASEAN MDD, NMPA (China), Taiwan, and 10+ 510(k) submissions. He has authored 40+ Clinical Evaluation Reports (CERs) for Class I–III devices in line with MEDDEV 2.7.1 Rev-4 and developed proposals for market access in the U.S., Europe, and APAC (including ASEAN, China, Taiwan, and Japan). He also prepared and implemented regulatory plans for new product development across 90+ countries through feasibility analysis and cross-functional coordination.

Intelligent Engineering

AI-Enabled Assessment of Carotid Stenosis using Dual-scan Ultrasound ViewsA Novel Approach using Deep Learning-Based Carotid Artery Quantification in Ischemic Strokes Evaluation using Transverse & Longitudinal Scans

Abstract

Introduction

Clinical Relevance of Carotid Stenosis Quantification

Literature Study

Methodology

Conclusion

References

About the Authors

About Cyient

EU AI Act Implementation Timeline

Use of AI in Healthcare