Tumor tissue pathology classification system and method based on adaptive proportional learning

Abstract

The invention discloses a tumor tissue pathology classification system and method based on adaptive proportional learning, and the method comprises the steps: firstly obtaining a plurality of pathological sections, carrying out the digital scanning, carrying out the manual marking of a scanned pathological image according to a classification task target category, and constructing a data set; segmenting the tissue foreground by using the difference distribution characteristics of RGB channels and gray values, and constructing a training data set of image blocks containing multi-stage magnification times; and finally, performing multi-stage amplification factor integration, combining the cross moisture function of each stage of amplification factor and the integration amplification factor to form a loss function, and achieving multi-amplification factor integration learning; and through adaptive proportion learning, performing dynamic adjustment on image global proportion labels and image block training weights which do not reach the lowest proportion, and therefore, the data utilization rate is increased, and rapid convergence is realized. In the pathological examination of daily tumor tissues, the detection rate is improved to the greatest extent on the basis of increasing extra workload as low as possible.

Description

technical field [0001] The present invention relates to the field of medical image processing and machine learning, in particular to a tumor histopathological classification system and method based on adaptive proportional learning. Background technique [0002] Pathological examination is the first step in cancer diagnosis and treatment, and the histopathological information provided is crucial for the judgment of cancer patients. However, due to the large shortage of pathologists, many pathological indicators cannot be fully and detailedly investigated in daily pathological examinations, and some rare tumor differentiations are only reported as moderately or poorly differentiated if they are not identified by pathologists. In the digestive system with a similar source of embryonic development, tumors in organs other than the liver may also show hepatic differentiation, without squamous epithelial distribution but may also show squamous cell carcinoma differentiation, and m...

AI Technical Summary

Problems solved by technology

[0005] 1. Manual reading is time-consuming and labor-intensive, and there are inter-individual and intra-individual differences. It is difficult to cover the entire pathological section and obtain finer-grained differentiation differences;

[0006] 2. Due to differences in sample preparation, staining standards, and scanning equipment, the data quality of digital pathology slide images varies greatly, and there are prone to non-parenchymal tissue areas including handwriting and edge artifacts. It is not suitable to remove these quality problems before For model training and direct application, quality control should be carried out before automatic identification to ensure data quality;

[0007] 3. The labeling cost of existing machine learning methods is too high: supervised learning requires a large number of pixel-level or image block-level labeling labels; it is difficult to have absolute complete labels for multi-instance learning, and there are too few training constraints, and the model effect is not good; full-scale pathology The objective scale label acquisition cost of the image is the same as that of supervised learning

[0008] 4. Existing models usually use only a single magnification or no specified magnification for modeling, and most tumor differentiation needs to combine morphological characteristics of different magnifications to obtain accurate judgments

Images