Research On Super-Resolution Reconstruction Technology Of Prostate Ultrasound Images Based On Deep Learning

With the rapid development of deep learning,computer-aided diagnosis of prostate ultrasound images has gradually been closely integrated with deep learning to help detect and treat diseases.However,there are problems with low resolution and unbalanced classification of prostate ultrasound images,and the current image classification algorithms have difficulty in exerting the performance of classification networks when the data volume is insufficient,resulting in low efficiency of intelligent assisted diagnosis of prostate ultrasound images.In order to solve these problems,this thesis proposes a deep learningbased method for super-resolution reconstruction of prostate ultrasound images,and applies it to the diagnosis of prostate cancer to improve the accuracy and reliability of computeraided diagnosis of prostate ultrasound images.Image super-resolution reconstruction technology and image classification technology in deep learning are used to process prostate ultrasound images to achieve assisted diagnosis of prostate ultrasound images.The main work of this thesis is as follows:(1)Research on super-resolution reconstruction technology of prostate ultrasound images fused with U-Net.Super-resolution reconstruction is a key method in data enhancement of prostate ultrasound images.The current super-resolution reconstruction algorithm uses a single-size convolution kernel to extract image features,which cannot effectively extract the feature information of the image,resulting in low feature utilization.To address this issue,the improved Cycle GAN is used for prostate ultrasound image superresolution reconstruction.U-Net network is fused to achieve multi-scale extraction of image features and enhance the reconstruction ability of the network.In the generator output of Cycle GAN,the output of U-Net is connected with the original input to retain more lowlevel features and detailed information.Fully exploiting the cycle loss of LR-SR-HR and HR-LR-SR and the adversarial properties of the discriminator,the generator is promoted to produce better perceptually consistent super-resolution reconstruction results.The experiment shows that compared with the classical Bicubic reconstruction method and the Cycle GAN reconstruction method applied to ultrasound images,the proposed method has good reconstruction results.(2)Research on super-resolution reconstruction technology of prostate ultrasound images based on EGDL-Cycle GAN.In the previous method,the image reconstructed by the Cycle GAN fusion U-Net model had large fluctuations in PSNR(Peak Signal-to-Noise Ratio)and SSIM(Structural Similarity)evaluation indicators,because the feature expression ability of the U-Net network is not sufficient.To address the above problems,based on the previous method,this thesis further studied the generator network and perceptual loss of Cycle GAN,and proposed an EGDL-Cycle GAN super-resolution reconstruction network for ultrasound images.The multi-scale feature extraction is achieved through the improved generator,and the full-scale skip connection between the encoder and the decoder is used to capture fine-grained details and coarse-grained semantics at the full scale.This effectively improves the performance of the generative network,resulting in better reconstruction results.This thesis proposes a new perceptual loss module to address the issue of information loss in traditional perceptual loss.The module utilizes residual structure to extract features deeply to obtain depth perception loss,and add it to the loss function for training the model,which can enable the model to learn the global and local differences between real and generated images,and pay more attention to the edge information and spatial information,and provide relevant spatial information feedback to the generator to improve the ability of the generator to perceive consistent SR.This method can enhance the prostate ultrasound image dataset,and provide rich images for the next step of intelligent assisted classification diagnosis of prostate cancer ultrasound images.The experimental results show that the EGDL-Cycle GAN method proposed in this thesis is superior to the classic image super-resolution reconstruction algorithm Bicubic,the perception-driven method SRGAN and the Cycle GAN method applied to ultrasound images in terms of PSNR,SSIM and visual effects.(3)Finally,this thesis designs an experiment to evaluate the performance impact of the original data set combined with the super-resolution reconstruction data set on the classification network.It is evaluated on the indicators of Accuracy,Precision,Recall,Specificity,and F1-score.The experimental results show that the classification after adding super-resolution reconstruction data network performance has been improved.