Immunohistochemical nuclear staining section cell positioning multi-domain co-adaptation training method

Abstract

The invention relates to an immunohistochemical nuclear staining section cell positioning multi-domain co-adaptation training method which is used for fully training a cell key point detection model under an only single-domain labeling data set. The method includes training the cell positioning model by adopting the source domain image and the target domain image, alternately inputting the sourcedomain image and the target domain image into an encoder to perform feature extraction, performing feature extraction on the source domain image to obtain a first feature, and performing feature extraction on the target domain image to obtain a second feature; inputting the first feature and the second feature into a discriminator for feature discrimination; when the loss function of the discriminator reaches a set condition, taking the extracted first feature and second feature as domain invariant features; alternately inputting the first feature and the second feature into a decoder for decoding, and performing activation operation to obtain a corresponding confidence map; in the training process, enabling the encoder and the decoder to perform parameter updating through continuous iteration; and when the number of training iterations reaches a specified number, ending the training.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a multi-domain co-adaptive training method for cell localization in immunohistochemical nuclear staining sections. Background technique [0002] Usually, the pathological aided diagnosis technology based on the deep learning method is trained on the digital pathological slides obtained after pre-processing by a single stain and scanner. However, due to the large variety of dyes and scanners on the market, in the same disease, digital pathology sections (especially immunohistochemical nuclear staining sections) produced by different institutions often have different distributions. Therefore, when the model trained in a single scene (source domain) is applied to multiple institutions (target domain), the performance of the model is often greatly affected due to noise distribution, data deviation, etc., and the test results are not satisfactory. However, the application scena...

AI Technical Summary

Problems solved by technology

And this has further caused that the performance of the key point detection model trained on the existing pathological image data set is greatly limited.

Therefore, how to fully train the cell key point detection model under the only single-domain labeled data set is a challenge

Images