Magnetic resonance reconstruction method of super-resolution convolutional neural network based on cavity convolution

Abstract

The invention discloses a magnetic resonance reconstruction method of a super-resolution convolutional neural network based on cavity convolution. The method comprises the following steps: step 1, selecting MRI as a training set; step 2, feature extraction: capturing features of a magnetic resonance image, performing convolution on an input image by using three hole convolution with different expansion rates to obtain three types of feature maps with different visual fields, adding residual learning, and directly combining output of an initial convolution layer with a result of hole convolution extraction to obtain three types of feature maps with different visual fields; the feature extraction capability of the network is further improved; and step 3, carrying out nonlinear mapping on the features extracted in the step 2. And step 4, reconstructing the mapped feature map. According to the method, the performance of MRI super-resolution reconstruction is improved, manual intervention or multi-stage calculation is not needed in practical application, and the effectiveness in MR image super-resolution reconstruction is greatly improved; the performance in the aspects of the evaluation index peak signal-to-noise ratio and the structural similarity is obviously improved, and the detail recovery of the image is also improved.

Description

technical field [0001] The invention relates to a magnetic resonance reconstruction method of a super-resolution convolutional neural network based on hole convolution, and belongs to the field of image processing. Background technique [0002] Magnetic resonance imaging (MRI) has been widely used in various clinical applications. Although powerful, MR imaging lasts relatively long and patients feel less comfortable. One of the tradeoffs to speed up the imaging process is to reduce image quality, using image post-processing techniques to improve MR image quality. [0003] In recent years, with the rapid development of deep learning, reconstruction algorithms based on super-resolution have been successfully applied in the field of medical imaging. A growing number of results tend to demonstrate the potential of deep neural networks for practical medical image applications. Therefore, the present invention proposes a MR reconstruction method based on a super-resolution conv...

AI Technical Summary

Problems solved by technology

Although powerful, MR imaging is relatively long in duration and less comfortable for the patient

Images