Research On Super-resolution Reconstruction Of Medical Ultrasound Signals And Images Using Generative Adversarial Network

B-mode ultrasound imaging is widely used in clinic.Ultrasound examination has the advantages that other medical imaging methods cannot replace,such as real-time,no radiation,and high cost performance.In real-time B-mode ultrasound imaging,the resolution of the image is limited by penetration depth,imaging time,and imaging equipment.To address this issue,this paper proposes to use super-resolution(SR)technology to increase the resolution of B-mode ultrasound images.In recent years,convolutional neural networks(CNN)have shown good performance in super-resolution reconstruction of natural images.But these methods have not been able to improve the image texture very well.The super-resolution generative adversarial network(SRGAN)solves this problem well,but it optimizes the super-resolution model in the image feature space and pays more attention to the human visual perception.The B-mode ultrasound image has the characteristics of speckle noise,which can be regarded as a grainy texture.For the super-resolution reconstruction of B-mode ultrasound images,the fineness of the speckle noise and the degree of similarity between the reconstruction result and the real image in pixel space are both important.This paper adopts and improves the SRGAN model.This paper mainly changed the network structure,loss function and training data set of SRGAN.Improved SRGAN is mainly used to increase the axial and lateral resolution of B-mode ultrasound images.The main goal of this work is to increase the resolution of the B-mode ultrasound image and make the reconstruction result as close as possible to the original high-resolution image.The experimental results show that,compared with the results obtained by commonly used SR method(such as the bicubic interpolation method used in MATLAB[Mathworks Inc.,MA,USA])and the SRGAN method,the high-resolution B-mode ultrasound image reconstructed by the improved SRGAN model in this paper has finer texture and sharper edges.The conclusion of this study is that the improved SRGAN in this paper can increase the resolution of the B-mode ultrasound image,improve the fineness of the speckle noise of the reconstructed B-mode ultrasound image,and reconstruct the result closer to the true high-resolution image.Motion displacement estimation is very important in quasi-static ultrasonic strain elastography(QUSE).It can be used to generate strain elastic maps to infer the hardness of tissues.In QUSE,motion tracking is performed between the pair of pre-and post-compression ultrasonic radio-frequency(RF)echo signals to obtain ultrasonically-tracked tissue displacements.The quality of the RF signal is closely related to the result of displacement estimation.The lateral resolution(perpendicular to the direction of the sound beam)of the ultrasonic RF signal is significantly lower than the resolution of the axial direction(parallel to the direction of the sound beam).So it is usually necessary to increase the lateral resolution of the RF signal.In this paper,before motion estimation,a super-resolution RF neural network(SRRFNN)based on generative adversarial network(GAN)is used to interpolate(upsample)the ultrasonic RF echo signals along the lateral direction to increase the lateral resolution of the RF signal.Our primary objective is to investigate the feasibility of using a GAN-based super-solution approach to improve lateral resolution in the RF data as a means of improving strain image quality in QUSE.Our preliminary experiments showed that,compared with axial strain elastograms obtained using the original ultrasound RF data,axial strain elastograms using ultrasound RF data up-sampled by the proposed SRRFNN model were improved.Based on the Wilcoxon rank-sum tests,such improvements were statistically significant(p<0.05)for large deformation(3-5%).In addition,from the perspective of improving the axial strain elasticity diagram,the SRRFNN model proposed in this paper is superior to the commonly used bicubic interpolation method.The conclusion of this study is that applying the proposed(SRRFNN)model can increase the lateral resolution of RF data and obtain good-quality strain elastography data in invivo tumor-bearing breast ultrasound data.