A multi-sequence magnetic resonance image registration method based on recurrent generative adversarial network

Abstract

A method for registration of multi-sequence magnetic resonance images based on cycle generation confrontation network, comprising the following steps: 1) training the original magnetic resonance images of sequence 1 and sequence 2 with CycleGAN, and outputting the magnetic resonance images of sequence 1 and sequence 2 Resonance generated images; 2) Perform single-modal registration on the generated images of the same sequence and the original images, and calculate the transformation matrix and the similarity measure between the two images of the same sequence; 3) Compare the similarity measures of the two sequences, and select the corresponding strategy to output the final transformation matrix; 4) use the final transformation matrix to transform the floating graph to obtain the final result graph. The invention has less dependence on registered samples, higher network trainability, stronger anti-interference ability and higher registration precision.

Description

technical field [0001] The invention relates to a multi-sequence magnetic resonance image registration method. Background technique [0002] Different weighted images are used in the acquisition of magnetic resonance images, resulting in different image representations of different sequences of magnetic resonance images. If the features are directly searched for the images, it is often impossible to match consistent features. Therefore, finding consistent features in the process of multi-sequence MRI image registration becomes a key issue for this type of registration, that is, converting MRI images of different sequences into images of the same sequence, and then using single-modality image registration method for registration. In the traditional machine learning image generation technology, a large amount of registered data is often required for training, such as the image generation technology based on structured random forest. The introduction of the method of generati...

AI Technical Summary

Problems solved by technology

[0004] In order to overcome the problem of insufficient anti-interference ability of existing generative adversarial networks in the conversion process of multi-sequence magnetic resonance images and the difficulty of network training, the present invention provides an easy-to-train network and strong anti-interference ability combined with cyclic generative adversarial A network-based transformation method for multi-sequence magnetic resonance image registration

Images