Unsupervised medical image segmentation method based on adversarial network

Abstract

The invention relates to an unsupervised medical image segmentation method based on an adversarial network, belongs to medical care informatics, and particularly relates to the technical field of medical image segmentation. According to the technical scheme, the method comprises the following steps: firstly, randomly generating or utilizing a third-party data set to obtain a group of auxiliary masks according to shape prior information, and sending the auxiliary masks and an unlabeled training image into a cyclic consistency adversarial network to generate binary masks; and a discriminator based on variational self-encoding and a generator correction module based on discriminator feedback are utilized to improve the quality of the binary mask. And after the binary mask of the training image is obtained, iterative training is performed by using a noise weighted Dice loss function, so that a final high-precision segmentation model can be obtained. According to the method, the problem that the convolutional neural network needs a large number of manual annotations in the training process of medical image segmentation can be solved, the problems of low performance, poor robustness and the like of an unsupervised segmentation method are solved, and the performance of an unsupervised medical image segmentation algorithm is effectively improved.

Description

technical field [0001] The invention belongs to medical care informatics, especially the technical field of medical image segmentation. Background technique [0002] Accurate segmentation of medical images is a very challenging task. Medical images have low contrast and blurred boundaries between different soft tissues, and different imaging modalities have huge differences. Images of different centers have large differences in contrast and resolution, making it very difficult to obtain accurate segmentation results. Traditional image segmentation methods such as level set, region growing algorithm, edge detection algorithm, etc. do not need to label images for training, and belong to unsupervised segmentation algorithms. They rely on artificially designed features and parameters, and are prone to over-segmentation and under-segmentation. Robust results are difficult to obtain with complex lesions. [0003] In recent years, deep learning and convolutional neural networks h...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to overcome the shortcomings of the existing deep learning algorithms in medical image segmentation that rely on a large number of manual annotations for training, and to solve the problems in the medical image segmentation method based on efficient learning of annotations, and propose an unsupervised medical image segmentation method based on adversarial networks. Image Segmentation Methods

At the same time, the invention overcomes the problems of generally low performance and complicated steps in the current unsupervised segmentation method, and enables the deep learning model to pay more attention to the global semantic information while paying attention to local details, so as to ensure the integrity of the segmentation results

Images