Medical image synthesis method based on generation countermeasure network

Abstract

The invention discloses a medical image synthesis method based on a generation countermeasure network, and relates to the field of image synthesis. A neural network generator branch of a healthy imageis synthesized through an infection focus image on the infection focus image to perform non-focalization on an infection focus area; a neural network generator branch of the infection focus image issynthesized through the healthy image on the healthy image to perform focalization on a certain area on the healthy image; a generation countermeasure loss function is constructed between a generationimage and a real image according to a generation countermeasure network model. The method has the advantages that cyclic consistency loss functions are constructed between the infection focus image and the infection focus image and healthy image correspondingly generated by two generators and between the infection focus image and the healthy image correspondingly generated by the two generators in order to stabilize the training of the neural network; fidelity loss functions are constructed on the infection focus image and the non-infection focus area of the correspondingly generated healthyimage in order to optimize the generated healthy image.

Description

technical field [0001] The invention relates to the field of image synthesis, in particular to a medical image synthesis method based on generating confrontation networks. Background technique [0002] By defocusing the lesion image, a one-to-one correspondence to the healthy image can be obtained, and this kind of medical image pair is impossible to obtain in reality. These one-to-one corresponding images can be used as training sample pairs for doctors to accelerate the learning process for doctors to distinguish lesion images from healthy images. The synthesized healthy images are of great help to computer-aided diagnosis, such as brain tumor segmentation and classification. Synthetic health images can also serve as augmented datasets, alleviating the lack of medical datasets. [0003] Focusing on healthy images can obtain one-to-one corresponding lesion images, and the synthesized lesion images can be used as an expansion of medical datasets. [0004] Since there are n...

AI Technical Summary

Problems solved by technology

[0004] Since there are no one-to-one corresponding sample pairs in real life, it is not reasonable to synthesize images in a supervised manner.

Xiaoren Chen published the article Unsupervised Detection of Lesions in Brain MRI using constrained adversarial auto-encoders at the International Conference on Medical Imaging with Deep Learning, 2018 to generate healthy images through variational self-encoders, but the resolution of healthy images generated by this method is low and does not generate high quality

Images