A method and device for segmenting abnormal tissues in H&E stained section images

Abstract

The invention discloses a method and device for segmenting abnormal tissues in H&E stained slice images. A feature cache module is constructed to store abnormal tissue coding features and normal tissue coding features, and the domain adaptive contrast loss is combined to guide the segmentation model to ignore the inter-sample intervals. The chromatographic domain difference, focusing on the feature difference between normal tissue samples and abnormal tissue samples, to solve the problem of poor prediction results caused by the different chromatographic domains between slices, so as to improve the prediction accuracy of the abnormal region of the segmentation model; at the same time, it is fully Use normal tissue samples for model training to reduce the model's over-reliance on expert annotation.

Description

technical field [0001] The invention belongs to the field of medical image processing, and in particular relates to a method and a device for segmenting abnormal tissue in H&E stained slice images. Background technique [0002] With the implementation of standardized diagnosis and the development of precision medicine, the requirements for pathological diagnosis reports are becoming more and more refined at this stage, and molecular pathology knowledge is updated at a fast pace. Pathologists are faced with the increase in the complexity of reports, the time required for reports and the amount of specimens. The challenge of growing contradictions. [0003] Due to their advantages of high throughput, homogeneity, and quantification, deep learning models are gradually becoming an effective way to solve the above problems. Deep learning networks such as U-Net, Res-UNet, deeplabv3+, etc. have been proposed specifically to perform various medical image segmentation tasks. These ...

AI Technical Summary

Problems solved by technology

[0006] The second difficulty is that digital scan images of tissue slices often have super-high resolution, which leads to extremely high time costs for experts to carry out fine labeling, which in turn makes training abnormal tissue sample data very scarce.

[0007] Patent document CN108447062A discloses a method for segmenting unconventional cells in pathological slices based on a multi-scale hybrid segmentation model. This method can realize the segmentation of unconventional cells based on a multi-scale hybrid segmentation model, reducing the frequent workload of pathologists, but Low prediction accuracy of the model due to failure to ignore differences between chromosomes during model training

[0008] Patent document CN109035269A discloses a method and system for segmenting lesion cells in cervical cytopathological slices. By introducing a semantic segmentation network constructed by multi-scale atrous convolution on the basis of a deep residual network and training the semantic segmentation model, the trained The semantic segmentation model segmented different types of diseased cells in the unit to be identified; combined with the morphological characteristics of pathological cells, the contour deformation model was established to further optimize the semantic segmentation results; The lesion category of the entire slice is predicted, and the semantic segmentation model and contour deformation model training process of this method does not ignore the difference between chromosomes, resulting in low prediction accuracy

Images