Automated flow-mediated dilation using neural network-based segmentation

Abstract Flow-Mediated Dilation (FMD) is a non-invasive technique used to assess endothelial function and, thereby, estimate atherosclerosis risk before the onset of its clinical manifestations. Cardiovascular disease remains the leading cause of death globally, responsible for approximately 17.9 million deaths annually, making early and accurate detection of cardiovascular risk critical. However, conventional FMD analysis relies on semi-automatic or manual measurement of arterial lumen diameter, which introduces observer variability and limits reproducibility. In this study, we propose an automated framework for brachial artery segmentation in FMD ultrasound videos based on deep learning. A dataset of 186 FMD examinations (920 manually annotated frames) was collected at the Clinical Research Center of the University of Campinas (UNICAMP) and split into 80% training and 20% validation sets. Three segmentation architectures were evaluated and compared: CMUNeXt, CMUNeXt-S, and CMUNeXt-L. The Deep Learning architecture CMUNeXt-L achieved the best performance among the models evaluated in this study, reaching an IoU of 0.9515. Furthermore, its results were superior to most traditional segmentation methods reported in the literature, including Rectangular Active Contour (Dice = 0.9335), Modified Affinity Propagation (Dice = 0.9213), and Cubic Splines Active Contour (Dice = 0.9170), while showing performance comparable to the Hough Transform combined with Canny edge detection approach (Dice = 0.9523). These results demonstrate that modern convolutional architectures with large-kernel inverted bottlenecks can substantially reduce inter-observer variability and accelerate the analysis of FMD examinations, supporting the use of FMD as a surrogate endpoint for clinical trials and its broader adoption in clinical practice.