Self-supervised ultrasound B-mode strain elastography using SMURF.
Authors
Affiliations (5)
Affiliations (5)
- Department of Electrical and Computer Engineering, UW-Madison, United States. Electronic address: [email protected].
- Department of Medical Physics, University of Wisconsin School of Medicine and Public Health, United States; Department of Medicine, University of Wisconsin School of Medicine and Public Health, 600 Highland Avenue, Madison, WI 53792, United States.
- Department of Medicine, University of Wisconsin School of Medicine and Public Health, 600 Highland Avenue, Madison, WI 53792, United States.
- Department of Neurological Surgery, UW-SMPH, 600 Highland Avenue, K4/8 CSC, Box 8660, Madison, WI 53792, United States.
- Department of Electrical and Computer Engineering, UW-Madison, United States; Department of Medical Physics, University of Wisconsin School of Medicine and Public Health, United States.
Abstract
Displacement tracking is an integral part of ultrasound strain elastography (USE) and various approaches have been proposed over the last three decades. However, limitations exist in the estimation of the lateral displacement field for several of these algorithms especially for in vivo imaging. Furthermore, most clinical systems output only B-mode image loops, necessitating the development of displacement tracking approaches on B-mode signals for wider clinical application of USE. In addition, since unsupervised deep-learning networks (DLN) require a large amount of unlabeled datasets, use of B-mode data can fill this void for clinical translation. In this study, we explore the use of a Self-Teaching Multi-Frame Unsupervised Recurrent All-Pairs Field Transform (SMURF), an optical flow approach for USE, focusing on estimation of strain tensor estimates in both the axial and lateral directions under an unsupervised setting using four-dimensional (4D) cost volumes on RF derived B-mode image loops. We retrain the RAFT network using baseline supervision on simulation dataset and the B-mode experimental dataset, and SMURF unsupervised training on B-mode experimental and in vivo datasets. We demonstrate that the fine-tuned RAFT model trained using unsupervised SMURF achieves comparable displacement and strain estimation accuracy and precision, especially in the lateral direction when compared to traditional window, and optimization-based methods for in-vivo datasets. The SMURF method also provides the fastest processing time when compared to other DLN with a 78% improvement over GPU based Lagrangian carotid strain imaging (LCSI) method. Our results demonstrate the potential of using self-supervised techniques for USE on B-mode datasets, unlocking the potential for real-time Lagrangian USE for clinical studies.