Convolutional Neural Network Based Speckle Tracking for Ultrasound Strain Elastography An Unsupervis

Accurate tracking of tissue motion is critically important for several ultrasound elastography methods including strain elastography and shear wave elastography. In this study, we investigate the feasibility of using a convolution neural network (CNN)-based speckle tracking method for elastographic applications. Additional data sets produced by finite element simulations were used to train an existing Flownet 2.0 model. After training, the improved network was evaluated using computer-simulated and tissue-mimicking phantoms, and in vivo breast ultrasound data. Our preliminary results were compared to a published 2D high-quality speckle tracking method by our group. Our preliminary results showed that the improved CNN outperformed the original Flownet 2.0. More specifically, in a tissue-mimicking phantom and one set of in vivo breast ultrasound data, the proposed CNN-based method achieved higher contrast-to-noise ratios (0.64 vs 1.05 and 1.16 vs. 1.40, respectively), as compared with the original Flownet 2.0 model. However, the improved CNN was still inferior to the coupled tracking algorithm. Currently, the inference process after the training of the proposed CNN can achieve approximately 60 frames/second for 2D speckle tracking under the NVIDIA TensorRT™ framework. Overall, we conclude that applying the proposed CNN-based speckle tracking method is feasible and good-quality strain elastography data can be obtained in TM phantoms and in vivo breast data. Our future work includes applying this technique to in vivo 3D whole breast ultrasound data.