Deep Multi-Task Attention Network for Automated, Reproducible Quantification of Carotid Plaque Burden in Ultrasound
Dawood Khan1, Kanwal Bano1, Hafiz Zia Ur Rehman2
1Department of Allied Health Sciences, Iqra University, Chak Shahzad Campus, Islamabad, Pakistan.
Ultrasound in Medicine & Biology
|June 5, 2026
Summary
This study introduces an attention-guided deep learning model for accurate carotid plaque quantification from ultrasound images, improving reproducibility and clinical assessment of cardiovascular risk.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular Disease
Background:
- Carotid atherosclerotic plaque burden is a key indicator of cerebrovascular and cardiovascular risk.
- Current quantitative assessment of plaque from ultrasound is operator-dependent and lacks standardization.
Purpose of the Study:
- To develop an automated, reproducible deep learning framework for quantifying carotid plaque area from B-mode ultrasound images.
- To improve the accuracy and reliability of carotid plaque burden assessment for clinical decision-making.
Main Methods:
- A deep multi-task attention-based learning framework integrating ResNet-50 and Convolutional Block Attention Modules (CBAM).
- Training and validation on 1,100 expert-annotated carotid ultrasound images.
- Joint learning of plaque presence and normalized plaque area within a unified architecture.
Main Results:
- Achieved a mean absolute error (MAE) of 0.001324 (2.44 mm²) and a Pearson correlation of 0.712 on the validation set.
- Demonstrated significant improvements in MAE (65.2%) over U-Net and (37.0%) over ResNet-50 regression.
- Bland-Altman analysis showed minimal bias (0.0003) and excellent reliability (ICC=0.85).
Conclusions:
- Attention-guided deep multi-task learning provides accurate and reproducible carotid plaque quantification.
- The model enables reliable detection of clinically significant plaque progression.
- This approach supports clinical translation by improving measurement reliability and agreement.
