Cardiac color Doppler from reduced packet sizes with deep learning

Abstract

Color Doppler echocardiography enables visualization of blood flow within the heart. However, the limited frame rate of color Doppler impedes the quantitative assessment of blood velocity throughout the cardiac cycle, compromising a comprehensive analysis of ventricular function. The formation of a color Doppler image involves ultrasound acquisitions that consist of approximately eight temporal frames, followed by clutter filtering to recover blood information, and subsequent Doppler velocity estimation. A potential solution to the low frame rate is reducing the number of temporal acquisitions for image reconstruction. However, conventional color Doppler processing methods for clutter filtering and Doppler velocity estimation are sensitive to reductions in temporal information. Recently, deep learning, and in particular convolutional neural networks, has demonstrated promising results in post-processing echocardiographic data for various applications. This thesis explores the use of deep learning models for color Doppler processing with a reduced number of temporal samples. We used a supervised learning method by creating simulated cardiac color Doppler images via a simulation pipeline that models the movement of both tissue and blood. We then explored attention-based U-Net models for clutter filtering, which outperformed typical high-pass filters. For Doppler velocity estimation from filtered signals, we proposed U-Net-based deep learning models combined with data augmentation strategies, which either matched or outperformed the state-of-the-art autocorrelator method, while effectively mitigating aliasing and noise. Overall, the proposed models demonstrated good generalization to in vivo data, despite being solely trained on in silico sequences. Finally, combining both methods yielded promising results on acquisitions with only three temporal samples. These findings demonstrate the potential of supervised deep learning methods for color Doppler processing using a reduced number of acquisitions.

Keywords

Echocardiography, Color Doppler, Clutter filtering, Doppler velocity estimation, Aliasing mitigation, Ultrasound simulations, Deep learning, Convolutional neural networks, Complex-valued neural networks, U-Net, ConvNeXt.

Jury: