Displacement estimation is a critical step in ultrasound elastography and failing to estimate displacement correctly can result in large errors in strain images. As conventional ultrasound elastography techniques suffer from decorrelation noise, they are prone to fail in estimating displacement between echo signals obtained during tissue deformations. This study proposes a novel elastography technique which addresses the decorrelation in estimating displacement field. We call our method GLUENet (GLobal Ultrasound Elastography Network) which uses deep Convolutional Neural Network (CNN) to get a coarse but robust time-delay estimation between two ultrasound images. This displacement is later used for formulating a nonlinear cost function which incorporates similarity of RF data intensity and prior information of estimated displacement . By optimizing this cost function, we calculate the finer displacement exploiting all the information of all the samples of RF data simultaneously. The coarse displacement estimate generated by CNN is substantially more robust than the Dynamic Programming (DP) technique used in GLUE for finding the coarse displacement estimates. Our results validate that GLUENet outperforms GLUE in simulation, phantom and in-vivo experiments.
Kibria, M. G., & Rivaz, H. (2018). GLUENet: Ultrasound elastography using convolutional neural network. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11042 LNCS, pp. 21–28). Springer Verlag. https://doi.org/10.1007/978-3-030-01045-4_3