Discussion
In this work, we have developed a deep learning approach for Doppler angle estimation. The clinical goal is to replace the repetitive and sometimes inconsistent manual Doppler angle adjustments. The chosen angle greatly affects the accuracy of blood flow velocity measurements, which in turn can have a major effect on the diagnosis and grading of arterial stenosis. In this study, a deep learning-based framework was used to automatically compute the Doppler angle from a set of pre-acquired ultrasound B-mode images. RMSE between the manual measurement of Doppler angle and the values estimated by the automated algorithm is suggested as a measure of the accuracy or the model. The results indicate that the RMSE was in the range of 3.96 to 9.27 degrees (for the different deep learning models evaluated) over the entire range of observed Doppler angles. However, in the range of Doppler angles 60 to 120 degrees, the mean difference was even smaller. This is attractive clinically because most ultrasound exams are performed with the carotid artery in the longitudinal orientation where the vessel is approximately perpendicular with the vertical axis. Hence, a clinical implementation of this algorithm would require that angles around 90 degrees are most accurately detected which has been demonstrated in this study.