An Enhanced Residual Cascaded Network Model for Undersampled Magnetic Resonance Imaging

Abstract

The invention discloses an enhanced residual cascade network model for magnetic resonance undersampling imaging. The invention provides the enhanced residual cascade network model, and the network takes densely connected recursive units as memory modules for learning bottom-layer feature information, and then takes dense connection as long memory connection among the memory modules for learning high-layer feature information; and compared with pure network cascading, such cascading method of local dense connection and global dense connection can better learn multi-level feature information andconstruct a deeper network structure. The cascaded deep neural network is constructed in combination with the residual network and the dense connection mode, so that better stability is provided fortraining of a deeper neural network, and the network performance is improved. In addition, a data consistency module and a high-frequency feature guiding module are added into the network for reinforcement, so that the confidence coefficient of a reconstruction result can be further improved, and the reconstruction effect of texture detail features can be further improved.

Description

technical field [0001] The invention relates to the fields of deep learning and medical imaging, and more specifically, relates to an enhanced residual cascade network model for magnetic resonance undersampling imaging. Background technique [0002] In clinical applications, medical imaging has become an indispensable diagnostic tool, and one of the important imaging methods is Magnetic Resonance Imaging (MRI). MRI is not only non-invasive, but also can acquire multiple imaging modalities with excellent contrast for resolving different anatomical features of disease. However, due to the long data acquisition time of MRI, the imaging target is required to be stationary, and the imaging price is high, which limits its wide application. In order to speed up MRI imaging, researchers have proposed a variety of fast sequential and parallel imaging techniques based on undersampling. By combining image post-processing methods, high-quality images that can be used for medical diagno...

AI Technical Summary

Problems solved by technology

Although many rapid imaging studies have demonstrated good clinical application prospects, current conventional MRI scans are still based on fully sampled Cartesian sequences, or use parallel imaging to accelerate acquisition

The undersampling reconstruction based on the traditional optimization algorithm still has limitations, which are mainly reflected in the following aspects: (1) The widely used sparse transformation is still too simple to deal with biological tissue images with complex structures

For example, although the TV-based sparse transformation can constrain the local information mutation of the reconstructed image, it will also introduce ladder artifacts; although the wavelet transform can enforce the isotropic information of the image, it will introduce block artifacts

(2) Nonlinear optimization algorithms usually require an iterative optimization process, resulting in too long reconstruction time, and the iteration may fall into local convergence

(3) The current reconstruction methods based on the optimization model generally need to set various optimization parameters, and inappropriate parameter settings will lead to over-constraints, resulting in unnatural reconstructed images, such as over-smoothing or residual under-sampling artifacts

Compared with popular models such as U-net, the CNN undersampling MRI reconstruction methods driven by the above models have higher mathematical resolution and shorter reconstruction time than traditional iterative optimization models; but the reconstruction results are not significantly superior. Especially when the sampling rate is only or even lower than 10%

Images