A comparative study of the state-of-the-art artificial intelligence models for automated plaque measurement in a carotid ultrasound Japanese cohort.
Authors
Affiliations (24)
Affiliations (24)
- Stroke Monitoring and Diagnostic Division, AtheroPoint™, Roseville, CA, USA.
- Department of Electronics and Communication Engineering, Bharati Vidyapeeth's College of Engineering, New Delhi, India.
- School of Artificial Intelligence, Sai University, Chennai, India.
- School of Biomedical Engineering, IIT (BHU), Varanasi, India.
- Department of Computer Science Engineering, Samrat Ashok Rajkiya Engineering College (SAREC), Uttar Pradesh, India.
- School of Computer Engineering, Kalinga Institute of Industrial Technology, Bhubaneswar, India.
- Department of Innovation, Global Biotechnologies, Roseville, CA, USA.
- Department of Radiobiology and Molecular Genetics, "VINČA" Institute of Nuclear Sciences - National Institute of the Republic of Serbia, University of Belgrade, Belgrade, Serbia.
- Division of Cardiology, Department of Medicine, Queen's University, Kingston, ON, Canada.
- Department of Biomedical and Molecular Sciences, Queen's University, Kingston, ON, Canada.
- Heart and Vascular Institute, Adventist Health St. Helena, St Helena, CA, USA.
- Department of Cardiology, Indraprastha APOLLO Hospitals, New Delhi, India.
- Invasive Cardiology Division, University of Szeged, Szeged, Hungary.
- Department of Research and Innovation, Uttaranchal Institute of Technology, Uttaranchal University, Dehradun, India.
- Department of CSE, IIIT, Bhubaneswar, India.
- Allergy, Clinical Immunology and Rheumatology Institute, Toronto, ON, Canada.
- University Hospital for Infectious Diseases, Zagreb, Croatia.
- Department of Radiology, Massachusetts General Hospital, Boston, MA, USA.
- Vascular Screening and Diagnostic Centre, University of Nicosia, Nicosia, Cyprus.
- Department of Rheumatology, University of Manchester, Dudley, UK.
- MV Diabetes Centre, Chennai, India.
- Department of Neurology, University of Cagliari, Cagliari, Italy.
- Department of ECE, Idaho State University, Pocatello, ID, USA.
- Symbiosis Institute of Technology, Nagpur Campus, Symbiosis International (Deemed University), Pune, India.
Abstract
Automated carotid intima-media thickness (cIMT) measurement from B-mode ultrasound is important for cardiovascular risk stratification. Direct comparisons of deep learning architectures on the same cohort under controlled conditions remain limited. This study aimed to compare the accuracy and computational cost of three deep learning architectures with the commercial AtheroEdge™ 2.0 (AE 2.0) system for carotid ultrasound analysis. This study evaluated three deep learning architectures: carotid non-attention double-stage convolutional neural network (cNonAtt-dsCNN), carotid attention-based single-stage UNet (cAtt-ssUNet), and carotid attention-based double-stage UNet (cAtt-dsUNet) against the commercial AE 2.0 system on 400 carotid ultrasound images from 200 Japanese patients. All deep learning models were trained using identical preprocessing, augmentation, hyperparameters (Adam optimizer, learning rate =1e-4, batch =32), and 5-fold cross-validation. Evaluation metrics included Dice similarity coefficient (DSC), Jaccard index, lumen-intima error (LIe), media-adventitia error (MAe), intima-media thickness error (IMTe), and computational cost. Under identical conditions, cAtt-dsUNet achieved the highest accuracy [DSC: 91.2%, Jaccard index: 85.5%, correlation coefficient (CC): 0.98, LIe: 0.031 mm, IMTe: 0.022 mm], reaching commercial-grade overlap and correlation but at ~2.5× longer training (20.2 <i>vs.</i> 8 min) and ~1.7× more graphics processing unit (GPU) memory than cNonAtt-dsCNN (DSC: 81.5%, CC: 0.92, LIe: 0.09 mm, IMTe: 0.124 mm). cAtt-ssUNet was intermediate (DSC: 88.5%, CC: 0.95, LIe: 0.07 mm). AE 2.0 retained the lowest boundary errors (LIe: 0.012 mm, IMTe: 0.011 mm) with 6 ms inference. A clear trade-off between accuracy-efficiency exists across the evaluated methods. cAtt-dsUNet is optimal for high-accuracy research settings with adequate GPU resources, while cNonAtt-dsCNN and AE 2.0 are preferable for speed-critical clinical screening. The single-cohort Japanese dataset limits generalizability; multi-center, multi-ethnic validation is recommended before clinical deployment.