Research On Generative Adversarial Networks Applied To Super Resolution Reconstruction Of Ultrasound Images

Ultrasound imaging is one of the most commonly used medical imaging methods at present.Itis real-time and convenient,and plays an important role in medicalimaging technology.However,ultrasound imaging has some drawbacks.When ultrasonictransmit in human tissues,there will be heterogeneous variation,such as blood flow,breathing and heart beating.These movements will interfere with imaging,and cause speckles and noise.These would interfere with doctor’s diagnosis.Therefore,improving the quality of ultrasound image is an important task.Super-resolution is an important means to improve image quality.By image enhancement,noise,speckle,blur can be removed and edge can be enhanced.In this paper,according to the characteristics of the ultrasound image blurring,we proposeusinga super-resolution reconstruction method combining with generator adversarial networkto enhance the ultrasound images.And improving the network to achieve better denoising effect.This paper mainly has the following work:1.Considering the pre-processing method of super-resolution which based learning using the bicubic downsampling to get the corresponding low-resolution image,but bicubic is based on the neighborhood information,and get a smooth results,discard the ground truth image informationwhich would cause poor effect of the resultsuper-resolution reconstruction.In order to further improve the reconstruction,we propose to use gaussian pyramid as pre-processing method,which can reserve more information of ground truth image,the result of the network and the ground truth similarity is higher,so we choose gaussian pyramid as the sampling method of the network.2.The super-resolution reconstruction was used in this method combination with the generative adversarial network.According to the characteristics of ultrasound imaging,the image was magnified four times with sub-pixel convolutional layer to obtain high-quality and clear large images.The residual block method is used to reserve the underlying information for better effect.Perception loss is added into the network,which is not only calculated at the pixel level,but also can make the visual effect of the generated results better.Comparing with Bicubic,SRCNN and ESPCN,the method not only has good visual effect,but also improves the PSNR and SSIM indexes.According to the feedback from clinicians,SRGAN produces vivid images with clear detail.3.Aiming at the speckle noise in ultrasound image,we improved the network structure,added gamma noise to the simulation images to simulate the speckle noise distribution of the ultrasound image,and image pairs are formed with the ground truth,composed with the simulation image.After network training,the denoised image is generated.Compared with BNLTV,SAR-BM3D and paper[91],the results of this method are clear in detail,and SSIM index is improved.Therefore,we know that the image super-resolution reconstruction combined with the generative adversarial networks can achieve good visual effect of ultrasound image;the improved GANs can remove speckle noise in ultrasound image and retain image details.