| Medical imaging technology is a technology that interacts the suspected lesions with electromagnetic fields and other media to display the tissue morphology of suspected lesions in the form of images.Doctors can intuitively verify their judgments and understand the extent of lesions based on medical images.Due to its unique advantages,magnetic resonance imaging has become more and more widely used in imaging examinations.High-resolution magnetic resonance imaging is the basis for doctors to make accurate diagnosis and subsequent precise treatment of diseases.However,high-quality magnetic resonance imaging is usually accompanied by long acquisition time,patient discomfort and cost of money.Therefore,super-resolution reconstruction technology has been proposed in recent years without changing the hardware configuration or increasing the scanning time.Super-resolution reconstruction technology refers to the reconstruction of unclear magnetic resonance images into high-resolution images that can be accurately diagnosed by doctors.This technology was first applied to the field of natural images.At present,it has been extended to the task of medical image processing and achieved good results.The existing medical super-resolution reconstruction algorithms only improve the performance by deepening the network or introducing general prior knowledge.To solve this problem,this thesis explores the unique prior knowledge of medical images,and studies the super-resolution reconstruction algorithm of medical images based on deep learning,in order to ensure the high resolution of the reconstructed image,so as to provide more detailed information for the pathological diagnosis of brain diseases and assist doctors in accurate diagnosis.In view of the problems existing in the current medical super-resolution.The main work of this thesis is summarized as follows:1.Knowledge distillation-based medical image super-resolution reconstruction algorithm:Aiming at the problem that the current method has never explored the ground-truth highresolution image as privileged information to assist network reconstruction,this thesis proposes a generalized distillation framework composed of teacher network and student network,which allows the teacher network to use privileged information during training,and transfer the important knowledge of the ground-truth image to the student network to assist the student network in super-resolution reconstruction;Aiming at the problem that the current methods only use the ground-truth image as the positive sample to guide the network training,this thesis proposes a contrast regularization method based on contrastive learning,which takes the information of the fuzzy image and the ground-truth image as the negative sample and the positive sample respectively,so as to ensure that the reconstructed image is closer to the ground-truth image.Experimental results show that the performance of the proposed algorithm is effectively improved based on the original model,and the reconstructed image has better visual effect.2.Segmentation prior-based medical image super-resolution reconstruction algorithm: Aiming at the problem that most methods do not consider the unique prior knowledge of medical images and only rely on the characteristics of low-resolution images to reconstruct highresolution images by deepening the network,this thesis proposes a medical image superresolution reconstruction network based on segmentation prior.The network effectively utilizes the positive influence of the prior knowledge of gray matter and white matter segmentation of magnetic resonance imaging on the super-resolution problem,and constructs a three-stage network model to restore the final high-resolution image.A large number of experimental results show that with the aid of prior knowledge of segmentation,the algorithm in this thesis is superior to the existing super-resolution reconstruction algorithms in PSNR and SSIM. |
