Intelligent Processing And Analysis Of Magnetic Resonance Images

Magnetic resonance imaging(MRI)is a non-invasive imaging method with high temporal and spatial resolution and no ionizing radiation,and has become the gold standard for clinical diagnosis of many diseases.MRI has been widely used in the diagnosis of various diseases and the visual guidance of targeted treatment.However,many physical and chemical factors are entangled throughout the magnetic resonance imaging procedure,making the processing and analysis of magnetic resonance image rather complicated,and many of them are even indeterminate problems that are difficult to solve.In recent years,as a breakthrough new method in the field of image analysis,deep learning has shown exciting prospects in medical image analysis.However,deep learning methods usually need a large number of sample data to support the training of the model.Due to the difficulty of magnetic resonance image data collection or label data acquisition,there are still great challenges in the successful application of deep learning in the field of magnetic resonance image processing and analysis.This paper aimed to design suitable deep learning algorithms for the problems with small sample size or no paired samples in the field of magnetic resonance imaging processing and analysis,and conducts in-depth research on typical imaging such as phase imaging,magnetic quantitative susceptibility mapping,and cardiac imaging.The contributions of this thesis are as follows:1)A PHU-Net model based on deep learning for MR phase unwrapping was proposed,which realizes the phase unwrapping without noise interference.One of the major challenges of deep learning to solve the problem of magnetic resonance phase unwrapping is that the real phase information is unavailable.To solve this problem,this paper designed a method to simulate the MR phase change patterns from natural images.Through comparative experiments,it is proved that the proposed data generation method was more suitable for training the magnetic resonance phase unwrapping model than the existing data generation methods,and solves the problem of insufficient sample size in model training;In the design of PHU-Net model,we proposed to replace the region growing in the traditional phase unwrapping method by integration and convolution,which improves the operation speed and accuracy of phase unwrapping;Additional noise filtering module was designed to enable PHUNet to realize phase unwrapping even when the signal-to-noise ratio was as low as 1.Finally,this method was directly applied to the water fat signal separation application,and a stable water fat separation was achieved for the dual-echo images.2)A TKD-net model for magnetic quantitative susceptibility mapping was proposed,which achieves similar results to supervised training without using real data.Quantifying magnetic susceptibility distribution from magnetic resonance phase images is a serious underdetermined problem in mathematics,and it is very difficult to obtain the true magnetic susceptibility information of tissues.Therefore,deep learning methods are limited by insufficient training samples when used to solve magnetic susceptibility quantification.This paper designed an unsupervised deep learning algorithm based on the intrinsic physical relationship between magnetic susceptibility and MR phase,utilizing the magnetic susceptibility distribution characteristics in k-space.Without using any real images for training,the method achieved similar performance to the supervised training method,and the structural similarity index(SSIM)reaches 0.986,even better than the supervised training method.The work provides a new idea for the current quantitative susceptibility mapping method.3)A deep learning transfer algorithm for magnetic resonance cardiac cine imaging was designed and successfully applied to predict the positive gene mutation of hypertrophic heart disease.Because the magnetic resonance cardiac cine imaging data is very difficult to obtain,the amount of available training data is small,which leads to the challenge of local extrema in the deep learning model.To address this problem,this paper combined the idea of transfer learning to design a deep learning scheme for hierarchical processing of each subtask.Each part of the scheme was designed to fit the objectives for different tasks,thereby reducing the number of training parameters of each part.The Area Under Curve of predicting the positive gene mutation reached 0.84,making the effective evaluation and prediction of patient gene positivity directly from magnetic resonance imaging data feasible.To summarize,this work aimed at the challenge of small samples or no present paired samples due to the principle of data acquisition or imaging in the field of magnetic resonance image processing and analysis,the corresponding model structure and training method are designed to provide new ideas for the intelligent development of deep learning magnetic resonance images.