Heart image registration system and method based on semi-supervised circulating GAN

Abstract

A heart image registration system and method base on semi-supervised cycle generation antagonistic network GAN are disclosed. That idea of the method is as follow: the cardiac computed tomography CT image and the transesophageal echocardiography TEE image are divided into blocks; The trained semi-supervised loop is used to generate the countermeasure network GAN, and the corresponding modal blockimage is obtained. Registering the block image to obtain the registration transformation matrix and the registered image. The block image registration transformation matrix with the largest normalizedmutual information value is migrated to the whole image to obtain the registration image of the whole image. The registration image and reference image are fused, and the fused image is displayed. The invention utilizes a semi-supervised cycle to generate an antagonistic network GAN, reduces the modal difference of a heart image, solves the registration problem of a large deformation image, improves the registration efficiency of the heart image, and improves the registration accuracy of the heart image by utilizing the local information after the heart image is divided into blocks.

Description

technical field [0001] The invention belongs to the technical field of image processing, and further relates to a cardiac image registration system and method based on a semi-supervised cyclic GAN (Generative Adversarial Networks) in the technical field of image registration. The invention can be used for registering the images of computed tomography CT (Computed Tomography) obtained before cardiac operation and images of transesophageal echocardiography (TEE) obtained during operation. Background technique [0002] Image registration refers to aligning two images into a common coordinate system so that changes between the two can be monitored. Registration based on feature points is one of the most commonly used image registration methods. Feature points need to be able to be distinguished from adjacent image points. If this condition is not met, it is impossible to uniquely match it with a corresponding point in another image. Therefore, a feature neighborhood should be...

AI Technical Summary

Problems solved by technology

The advantage of this method is to automatically determine the reference image and the region of interest, combined with the rigid body and non-rigid body registration method to achieve the global and local registration of four-dimensional myocardial perfusion MRI images in time series, but the shortcomings of this method are: , this method is to obtain reference images through heartbeat timing operations, and limit the registration image to one heartbeat cycle. When the images to be registered are different heartbeat cycles, the registration accuracy is low

However, the disadvantage of this method is that this method requires manual interaction during image segmentation, which makes the registration process cumbersome and affects the registration efficiency.

Images