Research On Ultrasound Image Detail Enhancement With Generative Adversarial Nets

The ultrasound signal passes through human bodies then its echo signal is processed by an ultrasonic instrument to form an ultrasound image.Doctors are able to diagnose by observing the ultrasound image.However,ultrasound probes with different frequency have their corresponding characteristics and applicable scenes.The images acquired by the high-frequency ultrasound probe have high resolution and fine details.While deeper organs cannot be observed using high-frequency probe as its display depth is not deep enough.Low-frequency ultrasonic probe can display large depth.However,the resolution of images are low and make it difficult for doctors to diagnosis.Doctors needs to choose suitable probe for diagnosis according to different clinical needs.Therefore,relying on the assistance of the computer to enhance the ultrasound image is of great significance to relieve doctor's stress.In this paper,two methods for image enhancement based on the Generative adversarial Nets are proposed helping doctors to diagnose with high-quality ultrasound images.In viewing the low-resolution ultrasound images,this paper use Cycle Generative adversarial Nets to map two features of low frequency and high-frequency ultrasound image data.The model can take low-frequency ultrasound images as input,and enhance image with high frequency details.In addition,for the problem of motion artifacts cause by the move of ultrasound probe during the acquisition,this paper uses Deblur Generative adversarial Nets(DeblurGANs)to make images clearer,with details enhanced.The above two image enhancement algorithms include the following three key parts:1.High-frequency detail enhancement using CycleGANs.The third chapter of this paper introduce CycleGANs to perform high-frequency detail enhancement on low-frequency ultrasound images.2.Improved W-CycleGANs.Aiming at the problem of original CycleGANs,a regular term using Wasserstein distance as a loss function is proposed to constrain training process so as to make better enhancement3.Reduce motion artifacts with DebluGANs.Learning the mapping of blurred image and clear image,the model can perform detail enhancement processing on the artifact generated by the probe motion during the ultrasonic image acquisition process.Quantitative indicators such as clarity,contrast and edge energy were used to evaluate the effects of different models.Finally,100 real IVUS image data of Guangzhou General Hospital in Guangzhou Military Region were used to verify the generalization of proposed network models.Results show that the improved W-CycleGANs model is twice as fast as the original CycleGANs model,and the three evaluation criteria are increased by 15.8%,11.4%and 46.6%,respectively.The deblurring effect of DeblurGANs is 48.6%and 19.9%higher than the benchmark method SRN.In summary,the two ultrasound image detail enhancement methods proposed in this paper are simple and effective.The W-CycleGANs can enrich the edge details of low-frequency images.The DeblurGANs can alleviate the motion artifacts generated during the image acquisition process.Both of them are able to enhance image details and improve the quality of images.