Semi-supervised medical image segmentation method and device based on dual-model interactive learning

Abstract

The invention provides a semi-supervised medical image segmentation method and a semi-supervised medical image segmentation device based on double-model interactive learning, which are assisted by a stability judgment strategy and are used for solving the false label quality problem of semi-supervised medical image segmentation. According to label data features, supervision constraints of cross entropy and DICE are introduced when label data knowledge is effectively learned. According to features of the false labels, adjoint variables are introduced in the method, and the adjoint variables are mainly used for relieving the influence of wrong false labels on the model learning process. And performing noise enhancement on the samples, and providing consistency loss based on a sample prediction result and a noise sample prediction result in a formal training stage. According to the learning of double models on label-free data, the method proposes a false label screening mechanism based on stability judgment, and realizes double-model interactive learning; the medical image segmentation method is simple and convenient to implement and flexible in means, and remarkable segmentation effect improvement is achieved on training data of the medical image.

Description

technical field [0001] The invention relates to the technical field of image segmentation, in particular to a semi-supervised medical image segmentation method and device based on dual-model interactive learning. Background technique [0002] In recent years, deep learning has achieved remarkable success in visual computer tasks such as image classification and image segmentation. Although fully supervised training deep learning models using finely labeled data have achieved very high performance in a variety of medical image segmentation tasks (such as neuron structure, polyp, liver, pancreas segmentation, etc.). But fully supervised training requires a large amount of pixel-level labeled data. In the real world, collecting accurate pixel-level labels for medical images requires medical experts to spend a lot of time, resulting in more expensive and time-consuming acquisition of medical image annotation data. Therefore, reducing the labeling cost of data, that is, buildin...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to provide a semi-supervised medical image segmentation method based on dual-model interacti

Images