A Deep Learning Approach to Resolve Velocity Aliasing in Blood Flow Velocity Measurement

The detection of hemodynamic parameters is crucial for early diagnosis of cardiovascular diseases. Traditional non-parametric methods for blood flow velocity measurement have drawbacks such as poor measurement precision and limitations imposed by the Nyquist limit. The recently presented one-dimensional convolutional neural network (1DCNN) method improves precision but suffers from significant fluctuations and suppression of reverse blood flow signals. To address these issues, this paper proposes a dual-branch convolutional neural network (DBCNN) method combined with singular value decomposition (SVD) for blood flow velocity measurement. Initially, SVD filtering was applied to simulated echo signals to remove substantial clutter. The filtered data was then processed by the DBCNN model for secondary filtering and blood flow velocity measurement. This model comprises two branches that estimate blood flow velocity magnitude and direction, enabling vector blood flow estimation and color imaging. Ultimately, the DBCNN model accurately estimated blood flow velocities within 1.2 times the Nyquist speed range, improving the performance of velocimetry and imaging. Compared to the SVD method using correlation algorithm and the 1DCNN model, the DBCNN model reduces the normalized root mean square error (NRMSE) for blood flow velocity profiles by an average of 32.61% and 19.7%, respectively.