Super-resolution Reconstruction Of Medical Ultrasonic Images Based On CNN

Because of the advantages of real-time and safety,ultrasound imaging is the most commonly used diagnostic method in clinical diagnosis.Ultrasound images often have low resolution,low contrast and speckle noise.The use of super resolution technology to improve their quality is becoming a research hotspot.FSRCNN is a good method of super resolution technology reported at present,but it is not ideal to reconstruct medical ultrasound image directly.In this thesis,an improved FSRCNN strategy for ultrasonic images is proposed,and a better result of ultrasonic image reconstruction is obtained.The average PSNR value is increased by about 1.3 dB,while the training parameters are reduced by 896.The main research work of this thesis is as follows:(1)The status of ultrasound image super-resolution technology at home and abroad is introduced,and the basic concepts of ultrasonic imaging and deep learning based super-resolution technology are introduced.(2)The effect of two deep learning methods of SRCNN and FSRCNN on ultrasonic image reconstruction is compared,and the FSRCNN,which has strong anti noise and short reconstruction time,is selected as the object of the algorithm.(3)FSRCNN is directly used to reconstruct medical ultrasound images,and reconstruction results are not ideal.This thesis uses standard image library to build a simulation model of FSRCNN training and testing.By adding different intensity noise to all the images in the image library,the influence of the noise intensity on the quality of the FSRCNN reconstructed image is analyzed.It is found that the FSRCNN model has the requirement for the quality of the reconstructed image and the training set image.This thesis uses the blind deconvolution algorithm of Rob to preprocess the original ultrasonic image and the reconstructed ultrasonic image.The clarity of texture details is improved significantly,and the image quality is increased by about 7dB to achieve the above requirements.The weights obtained from pre-processing training set are used to reconstruct ultrasound images,and the reconstruction quality is increased by about 0.54 dB.(4)In order to further improve the quality of reconstruction,This thesis takes the medical ultrasonic image as the research object,establish the simulation model of FSRCNN,optimize the number of convolution kernel number,convolution kernel and deconvolution convolution kernel in FSRCNN,and find that when d=56,b=3* 3,s=9*9,the quality of the reconstruction of ultrasonic images is the most.OK,and the training parameters were reduced by 896.(5)In this thesis,FSRCNN algorithm is improved for medical ultrasound images: first,the ultrasonic image is input,and the optimized FSRCNN is trained to get the weight parameters;then the ultrasonic image is reconstructed with the weight parameters obtained in the last step,and the final reconstructed ultrasonic image is obtained.Finally,using ultrasonic images from the ultrasonic imaging Institute of Medical University Of Chongqing,we verify the reconstruction strategy of CNN and improve the quality of ultrasonic image.