Research On The Method Of Reconstructing Ultrasound Radio Frequency Signals Into Images Based On Deep Learning

The ultrasonic radio frequency signal is the signal obtained after pre-amplification,A / D conversion and beam synthesis of the echo signal obtained from the transducer,because it contains relatively complete sound field and tissue interaction information.Therefore,many ultrasound devices use radio frequency signals as the original signals for ultrasound image reconstruction.The signal is used as the original signal for ultrasound image reconstruction.Since many steps are involved in the traditional ultrasound imaging process,many parameters need to be set for each step,especially the image post-processing part.The ultrasound imaging systems of different manufacturers use different methods,and their imaging styles and quality are different and are based on your own experience to adjust the parameters.Therefore,it is of great significance to propose a new general ultrasound imaging method,which can not only replace the steps that require manual parameter adjustment in the reconstruction process,but also achieve the traditional ultrasound imaging level.In this study,the goal is to achieve a complete reconstruction process from the RF signal to the ultrasound grayscale image.This paper uses a depth method to design an end-to-end reconstruction model,which inputs the envelope information of the RF signal and outputs the ultrasound image.According to the characteristics of the radio frequency signal,this paper preprocesses the radio frequency signal to obtain the signal envelope,and then uses the condition based on supervised learning to generate an adversarial network to learn the mapping process from the envelope image to the ultrasound image.Then use the trained network as the reconstruction model.The radio frequency data used for network training includes Field II simulation data,phantom data and human body data,in which the phantom and human body data are collected on two ultrasound equipment of different brands.The purpose is to verify that the reconstruction model has reconstruction capabilities on different ultrasound equipment.At the same time,this article uses the method of transfer learning,the reconstruction model trained on the simulation data is transferred to the phantom data for training,and then the phantom model is transferred to the human body data.At the same time,the simulation,phantom and human body data test sets are used to test the respective trained reconstruction models,and the model trained on the human body data is tested on the external data set.This paper calculates the similarit y between the image re constructed by the network and the image reconstructed by the tradit ional method and the B-mode image collected by the ultrasound equipment.The structural similarity(SSIM)of the reconstructed network on the phantom,human body and external data is 0.82,0.63 0.55;traditional methods are 0.47,0.32,0.30.The SSIM obtained on the two different devices with the reconstructed network has an average difference of 0.07 on each data,while the difference between the traditional methods is 0.11.From the results,the reconstruction method based on deep learning proposed in this paper is superior to the traditional reconstruction method in phantom,human body or external data.The image quality reconstructed on phantom data is very close to that of B-mode images.On the human body and external data,the structural outline informatio n and spatial position information of the image can be reconstructed well,and the reconstruction qualit y can be achieved on both ultrasound equipment.Therefore,the reconstruction method based on deep learning proposed in this paper can replace part of the traditional reconstruction method steps and basically reach the imaging level o f ultrasound equipment.