Fully automated dual-wall intima-media thickness measurement in carotid ultrasound using deep learning-based segmentation models

Carotid artery intima–media thickness (IMT) is an ultrasound-derived biomarker of arterial wall structure and vascular remodeling. Although its role in routine cardiovascular risk stratification has been de-emphasized in contemporary guidelines, IMT remains widely used in vascular imaging research. Conventional carotid ultrasound measurements are limited to far-wall IMT due to acoustic shadowing and interface artifacts, potentially underestimating the carotid arterial wall morphology by missing near-wall pathology. To develop and technically evaluate a deep learning algorithm for automated measurement of both near-wall and far-wall IMT from carotid ultrasound images, and to assess segmentation accuracy and measurement reproducibility across multiple model architectures. A total of 641 carotid B-mode ultrasound images were obtained from the Kangbuk Samsung Health Study cohort. Ten segmentation architectures and fifteen backbone networks were systematically evaluated. Based on the predefined selection criteria and five-fold cross-validation results, LinkNet with a ResNet-152 backbone was selected for subsequent evaluation. External validation was performed using 128 images from the publicly available Carotid Ultrasound Boundary Study multicenter dataset. Segmentation performance was assessed using Intersection over Union (IoU) and Dice similarity coefficient (DSC); measurement accuracy was evaluated using mean absolute error (MAE) and mean squared error (MSE). Among ten evaluated architectures, LinkNet achieved the highest mean IoU (0.836) and mean DSC (0.910) using ResNet-50 as a standard backbone. With ResNet-152, LinkNet achieved the highest internal validation performance, with a mean IoU of 0.849 and mean DSC of 0.918 (far wall: IoU 0.879, DSC 0.936; near wall: IoU 0.820, DSC 0.901). Near-wall IMT measurement error was 0.118 mm MAE. On external validation, mean IoU was 0.803 and mean DSC was 0.890, with a near-wall MAE of 0.116 mm. Performance was consistently lower for the near wall compared to the far wall across all models, reflecting the inherent acoustic imaging challenges in near-wall delineation. The proposed deep learning pipeline demonstrated the technical feasibility of automated segmentation and measurement of both near-wall and far-wall carotid IMT. Although encouraging performance was observed in both the internal and external datasets, near-wall segmentation remained more challenging than far-wall segmentation. Given the limited number of participants and the restricted external validation subset, these findings should be regarded as proof-of-concept results that warrant further validation in larger and more heterogeneous cohorts.