Research On MR Image Reconstruction Algorithm Based On Deep Learnin

Magnetic Resonance(MR)imaging is a powerful medical imaging method that does not involve ionizing radiation and can be used as a medical diagnostic tool.However,routine clinical MR scans are time-consuming and expensive,and can cause discomfort to patients.Currently,MR data acquisition can be accelerated through undersampling,but different undersampling methods can produce different types of artifacts and noise during the imaging process,leading to poor reconstruction results that can affect the diagnostic outcome and potentially harm the patient.In recent years,deep learning-based image reconstruction algorithms have made significant progress,with convolutional neural networks’ powerful image feature extraction ability widely applied in the field of image processing.Based on this,this study uses deep learning technology to investigate the MR image reconstruction problem under different undersampling strategies and proposes different solutions in the image domain and Kspace domain to more comprehensively address the problem.(1)Aiming at the problems such as loss of key information and poor robustness in the traditional Cartesian undersampling reconstruction,an image domain based deep learning reconstruction model: Unsupervised Alternate Iterative Optimization Convolutional Neural Networks(UA-CNN)was proposed.Based on the traditional convolutional neural network,UA-CNN integrates the plug-and-play iterative optimization algorithm and the unsupervised feature loss module to solve existing algorithm problems such as key information loss,weak model fitting capability,and weak robustness to noise by effectively constraining the model.Experimental results show that UA-CNN outperforms other algorithms in various evaluation metrics and generates reconstruction images with clearer and more coherent image details.(2)To address the issue of massive data volume and uneven K-space distribution in existing reconstruction algorithms,a data-driven cascaded neural network model is proposed.This model introduces a data consistency layer and convolutional neural network modules in different domains(K-space and image domain)through a cascading approach,enabling the model to learn more image features with a small amount of data and solve problems such as missing feature selection mechanism and uneven K-space distribution in the original model training process.The research indicates that the proposed model has good generalization performance and high noise tolerance,and the method implemented on GPU is more than 75 times faster than CPU algorithms.In summary,experimental results of reconstruction under different undersampling strategies demonstrate that UA-CNN can effectively remove artifacts and noise while preserving image texture details and edge information,and the data-driven cascaded neural network can reconstruct high-quality images with good convergence performance even with a small amount of data.This study validates the superiority of deep learning-based algorithms in MR image reconstruction research and provides an efficient and fast solution for MR image reconstruction.