An unsupervised learning approach to ultrasound strain elastography with spatio-temporal consistency

Abstract

Quasi-static ultrasound elastography (USE) is an imaging modality that measures deformation (i.e. strain) of tissue in response to an applied mechanical force. In USE, the strain modulus is traditionally obtained by deriving the displacement field estimated between a pair of radio-frequency data. In this work we propose a recurrent network architecture with convolutional long-short-term memory decoder blocks to improve displacement estimation and spatio-temporal continuity between time series ultrasound frames. The network is trained in an unsupervised way, by optimising a similarity metric between the reference and compressed image. Our training loss is also composed of a regularisation term that preserves displacement continuity by directly optimising the strain smoothness, and a temporal continuity term that enforces consistency between successive strain predictions. In addition, we propose an open-access in vivo database for quasi-static USE, which consists of radio-frequency data sequences captured on the arm of a human volunteer. Our results from numerical simulation and in vivo data suggest that our recurrent neural network can account for larger deformations, as compared with two other feed-forward neural networks. In all experiments, our recurrent network outperformed the state-of-the-art for both learning-based and optimisation-based methods, in terms of elastographic signal-to-noise ratio, strain consistency, and image similarity. Finally, our open-source code provides a 3D-slicer visualisation module that can be used to process ultrasound RF frames in real-time, at a rate of up to 20 frames per second, using a standard GPU.

An unsupervised learning-based shear wave tracking method for ultrasound elastography

“…However, the difference in terms of displacement accuracy between ReUSENet and USENet is not as important as in our previous work on quasi-static elastography, where we showed that feed-forward networks failed to estimate large range quasi-static deformations. 7 In shear wave imaging, the displacement generated is in the order of a few micrometres, and the ultrasound time series are acquired with a high temporal sampling rate (about 5-10 ms). Therefore, our results suggest that both feed-forward and recurrent neural networks can efficiently track shear wave displacements, and that the use of convLSTM units is better suited for ultrasound time series that exhibit larger deformations.…”

“…The architecture of both neural networks used in this study has been described in details in our previous works, and are named USENet and ReUSENet for '(Recurrent) Ultrasound Elastography Network'. 7 They both consist in an encoder-decoder neural network with skip connections. The encoder part is the same for both networks, and is composed of four down-sampling ResNet blocks.…”

“…6 In this work, we investigated the use of CNNs to improve displacement estimation accuracy for tracking shear wave propagation, directly from RF ultrasound data. We considered two different network architectures proposed previously in strain elastography, 7 and compared them with a conventional normalised cross-correlation method. The results presented were obtained from simulated ultrasound RF data.…”

Ultrasound Shear Wave Elasticity Imaging With Spatio-Temporal Deep Learning

“…Recently, deep learning methods have gained popularity in strain elastography ( [17], [18], [19], [20]) and SWE-imaging ( [21], [22], [23]). These methods allow estimates without intensive preprocessing of the data, manual tuning and do not rely on feature extraction, e.g., the shear wave velocity for elasticity estimation.…”

Physically Inspired Constraint for Unsupervised Regularized Ultrasound Elastography

“…They have been successfully adopted for USE by modifying the architectures to handle high-frequency radio-frequency (RF) data [14,6]. Unsupervised and semi-supervised techniques have also been employed to train the networks using real US data without requiring the ground truth displacements [13,1,2,15]. They employed prior knowledge of displacement continuity in the form of the first-and second-derivative regularizers.…”