A Super-resolution Method Of Quantitative Magnetic Resonance Imaging Based On Deep-learning

Quantitative magnetic resonance imaging(q MRI)is an essential technique of medical imaging clinically.However,acquisition of such high signal-to-noise ratio and high-resolution(HR)data is time consuming,and could lead to motion artifacts.Therefore,there is a trade-off for MRI between resolution and scanning time.Super-resolution MRI achieved with deep learning is a promising approach,which can recover superior HR images from their low-resolution(LR)counterparts compared with polynomial interpolation methods.State-of-the-art SR methods are mostly supervised methods,which require external training data consisting of specific HR-LR pairs.However,such training data in pairs are often unavailable for MRI due to scanner limitations.Finally,current learning-based SR methods for MR images have not considered the quantitative conditions,therefore,the estimated quantitative map from a set of weighted images after SR reconstruction is changed and inaccurate.In this paper,a deep-learning based super-resolution method for quantitative magnetic resonance imaging(MRI)is proposed.This method is an unsupervised learning method,which makes use of the self-similarity of images combined with the learning ability of the deep network,and does not need to obtain additional paired LR-HR quantitative images to train the network,only the low-resolution images collected can get better results.At the same time,because the existing learning-based SR method does not take into account the quantitative conditions of magnetic resonance,this paper takes 1magnetic resonance quantitative image as an example,the 1exponential decay model is introduced to constrain the proposed super-resolution model to obtain accurate quantitative parameter maps.The experimental results show that,compared with the current mainstream SR method,it has a better performance both in the comparison of objective indicators and in the subjective evaluation.