A Bayesian and Minimum Variance Technique for Arterial Lumen Segmentation in Ultrasound Imaging

Cardiovascular diseases (CVDs) are the worldwide leading cause of deaths. Based on ultrasound, primary assessment of CVDs is measurement of carotid intima-media thickness and brachial endothelial function. In this work we propose improvements to a state of the art automatic methodology for arterial lumen detection, based on graphs and edge detection, fundamental for cited tests. We propose a bayesian approach for segmenting the graph minimum spanning tree created with points between edges. Lumen is located applying three criteria on segmented trajectories: length, darkness, and our proposal, minimum variance. In 294 sonograms having manually established measurements, from a 1,104–sonogram set, mean and standard deviation error in brachial near wall detection was 14.6 μm and 17.0 μm, respectively. For far wall they were 15.1 μm and 14.5 μm, respectively. Our methodology maintains superior performance to results in recent literature that the original methodology presents, but surpasses it in overall accuracy.