Cardiovascular diseases remain one of the leading causes of global mortality. Flow-mediated dilation of the brachial artery, assessed by ultrasonography, is a non-invasive method to evaluate endothelial function; however, manual analysis is time-consuming and prone to variability. This study proposes the automatic segmentation of the brachial artery lumen using computer vision techniques based on deep learning. We investigated three segmentation architectures: UNet++, MultiResUNet, and TransUNet, applied to ultrasound images enhanced with contrast adjustment through CLAHE. Model training was conducted using K-fold cross-validation, with performance assessed through established metrics such as Dice, IoU, and Accuracy, while employing the Focal Tversky Loss function, suitable for imbalanced scenarios. Among the evaluated approaches, UNet++ achieved the best results, with Dice of 0.950, IoU of 0.905, and Accuracy of 0.980. These findings demonstrate the potential of deep segmentation models to reduce the subjectivity of clinical analysis and improve diagnostic accuracy in ultrasound-based vascular assessments.
Silva et al. (Tue,) studied this question.