Research On Ultrasonic Speckle Motion Tracking Algorithm And Elastic Imaging Based On Optical Flow Networ

Accurate tracking of tissue motion is critically important for several ultrasound elastography methods including strain elastography,acoustic radiation impulse force imaging and shear wave elastography.In this study,we investigate the feasibility of using three published convolution neural network(CNN)models built for optical flow by the computer vision community for breast ultrasound strain elastography.Elastographic data sets produced by finite element and ultrasound simulations were used to retrain three published CNN models(transfer learning):FlowNet-CSS,PWC-Net,and LiteFlowNet.After retraining,the three improved CNN models were evaluated using computer-simulated and tissue-mimicking phantoms,and in vivo breast ultrasound data.CNN-based tracking results were compared to two published 2D speckle tracking methods:Coupled and GLUE methods.In order to achieve high-precision tracking of two-dimensional ultrasound data by optical flow network and improve the sensitivity of the network to real ultrasound data,the following two aspects of research work are carried out in this paper.(1)Select CNN model for ultrasonic data.Because most of the well-known optical flow networks that have been published are applicable to ordinary optical images such as vehicle motion,and rarely involve the field of ultrasonic elastic imaging.In order to better carry out migration training for these models,this paper chooses flownet2.0,PwC-Net and LiteFlowNet as the research basis based on the comprehensive consideration of training time,network performance and others.But there are many versions of these networks.Therefore,several real breast ultrasound data are used to test the original optical flow network of various versions.Based on the analysis of the preliminary visualization results,flownet-CSS,PWC-kitti and LiteFlowNet-sintel were selected for transfer training.(2)According to the structure of the model,different migration training methods are designed to train CNN model from coarse to fine or from simple to complex.The training process of all models is divided into stages.First,the model is initially trained by using the simple simulation data of computer simulation,and then the model is refined by using the complex breast ultrasound simulation data.When the model converges,combined with visual analysis and quantitative evaluation,Using various ultrasound data to analyze the model in breast ultrasound elastography.Our preliminary results showed that,after retraining,in simulated uniform phantoms,all three CNN models significantly outperformed the coupled tracking method.Retraining was effective for in vivo cases as well.The mean CNR values of axial strain using those three original models among 31 in vivo cases were 0.88(FlowNet),0.82(PWC-Net)and 0.58(LiteFlowNet),respectively,whereas the mean CNR values were improved to 1.15(Retrained-FlowN et),1.43(Retrained-PWC-Net)and 1.02(Retrained-LiteFlowNet),respectively.Overall,based on the Wilcoxon rank sum tests,the improvements due to retraining were statistically significant(p<0.01)for the PWC-Net and LiteFlowNet models.We also found that the PWC-Net model was the best neural network model for data investigated and its overall performance was on par with the coupled tracking method.The PWC-Net model was also able to achieve approximately 45 frames/second for 2D speckle tracking for data investigated.