Pathological image dyeing style normalization method and device

Abstract

The embodiment of the invention provides a pathological image dyeing style normalization method and device, and the method comprises the steps: inputting a first pathological image to be subjected todyeing normalization into a dyeing style normalization model which is located in a central server and completes the training, and outputting a dyeing style between the dyeing styles of all data centers, while not changing the second pathological image of the image morphological characteristics of the first pathological image, so that the normalization of the dyeing style of the pathological imageis realized. The output of the color style normalization model is used as the input of the color style discrimination model in each data center, and forms a generative adversarial network with the color style discrimination model in each data center to carry out training. According to the embodiment of the invention, the training of the dyeing style normalization model does not need to obtain theoriginal image of each data center, so that the privacy of the data is ensured, and the problem that the dyeing style deviation is easy to occur because the template image needs to be selected for dyeing by using the priori knowledge of the pathological image in the traditional method is avoided.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method and device for normalizing the coloring style of pathological images. Background technique [0002] In federated learning, data from different data centers need to be fused to jointly train a robust and powerful deep learning model. This machine learning method is especially important for medical images with small data volumes and difficult labeling. However, the staining style of pathological images is quite different due to the different staining techniques, staining conditions and equipment of each pathologist, which affects the final identification and classification. [0003] At present, the normalization methods for pathological image staining in the prior art include: ①Collect image samples from multiple data centers together, use prior knowledge to select template images, and then use staining mathematical algorithms to combine them with The im...

AI Technical Summary

Problems solved by technology

[0004] However, in the above two schemes, the server must obtain all the original image data of each data center when training the model, which cannot meet the requirements of federated learning for the confidentiality of image information in each data center

And for scheme ①, there will often be errors in coloring style deviation

For scheme ②, it is often limited to two data sets, and after the style of the data set selected as the template changes, the accuracy rate will drop significantly, which is not suitable for training and testing data sets containing multiple data The case of pathological samples with multiple staining styles in the center is also not suitable for the setting of multiple data centers in federated learning

Images