Preliminary investigation on clutter filtering based on deep learning

Abstract

In recent years, singular value decomposition (SVD)-based clutter filters have received widespread attention in ultrasound flow imaging owing to their high performance over traditional clutter filters in suppressing clutter signals. The excellent performance of the SVD clutter filter depends on its adaptive nature. The SVD clutter filter adaptively rejects echoes from slowly moving clutters, allowing visualization of echoes from blood cells. Owing to this property, the SVD filter works well throughout a cardiac cycle. Recently, deep neural networks have been used for a variety of tasks. The adaptive nature of deep neural networks would be beneficial for clutter filtering in ultrasonic blood flow imaging. In the present study, we conducted a preliminary study on clutter filtering using a long short-term memory neural network. Experimental results suggested that the proposed deep-learning clutter filter achieved a comparable performance than SVD one in terms of contrast values.

Robust adversarial learning model to segment non-speckle regions in blood flow echo

“…Hence, additional features and other deep learning models, e.g. long-short term memory 17) as well as the developed U-Net model, would be required for the feedback of frame data. In addition, since there is system noise in in vivo data, we need to consider plane wave compounding and ensemble averaging of several frames to reduce noise.…”

“…Our previous studies have also investigated the strategy using convolutional neural network (CNN) to enhance the contrast between blood flow (lumen) and surrounding tissues. 17,18) CNNs can extract and process complex features that consider the spatial information of images by alternately passing through convolutional, pooling, and fully connected layers. 19) Several auto-encoder type models have been proposed to improve CNNs for segmentation.…”

Ultrasound Signal Processing: From Models to Deep Learning

“…To that end, they used SVD-processed in vivo and in vitro images as targets. Similarly, Wang et al (Wang et al, 2021) aimed at replacing SVD thresholding, but rather than a 3-D CNN, the authors adopted a 2-D (spatial) CNN, aggregating temporal information in the feature space through a recurrent neural network.…”

Ultrasound Signal Processing: From Models to Deep Learning

“…To that end, they use SVD processed in-vivo and in-vitro images as targets. Similarly, Wang et al aim at replacing SVD thresholding(Wang et al, 2021), but rather than a 3D CNN, the authors adopt a 2D (spatial) CNN, aggregating temporal information in the feature space through a recurrent neural network.Tabassian et al (2019) use a deep 3D CNN (2D + time) to suppress clutter and reverberation artifacts that plague echocardiographic imaging. Their deep network was trained on realistic simulations of echocardiographic exams, with simulated superimposed artifacts.…”