Medical image automatic labeling method and system based on small sample segmentation

Abstract

The invention requests to protect a medical image automatic labeling method and system based on small sample segmentation, and aims to solve the problems that a large amount of labeled data is needed for medical image segmentation, the process of labeling new data is tedious and single, but a large amount of manual labeling work is needed, and the cost of a data set is increased. On the basis of a small sample segmentation technology, an automatic annotation network structure Siamese-DCNet is provided, a double-branch structure is utilized, a query branch and a support branch are included, features of an unannotated image and features of an annotated image are extracted preliminarily, and by means of a result obtained by the double branches and in combination with known annotations, unimportant information except the annotation is removed; a preliminary annotation is predicted by calculating cosine similarity and is input into an iterative optimization module, and a final annotation result is obtained through refining of several iterations. According to the method, automatic annotation of all other images in the same scene can be realized only by a small number of images with annotations.

Description

technical field [0001] The invention belongs to the technical fields of deep learning, image processing, medical image segmentation, and automatic labeling, and in particular relates to an automatic labeling method for medical images based on small sample segmentation. Background technique [0002] In the field of medical images, the annotation results of medical images can assist medical workers to make reasonable judgments on patients' conditions and formulate corresponding diagnostic methods. In recent years, with the widespread application of deep learning image segmentation technology in many computer vision applications (eg, autonomous driving, medical imaging, remote sensing technology), more and more image data needs to be used to train deep learning models. However, since the objects in medical images are of different sizes, poses and shapes, and the boundaries are not obvious, the labeling of images is a very time-consuming and laborious work. In addition, in orde...

AI Technical Summary

Problems solved by technology

And when the number of training image samples is limited, the nearest neighbor method may not work well

Images