Recognition Of Cell In Pathological Image Based On Deep Learning And MIL

Pathological examination based on histological sections is the solid standard for tumor diagnosis.Pathological tissue sections are usually scanned by Hematoxylin-Eosin（H＆E）staining and Whole Slide Imaging（WSI）technology,which has ultra-high resolution and can fully display the tissue The microscopic expression of cells and cell microenvironment provides an important basis for tumor diagnosis.However,this also brings a key problem to a class of computer-aided diagnosis technologies represented by deep learning: how to identify cell-level microscopic features from ultra-high-resolution images for auxiliary diagnosis of cancer,so as to reduce the repeated identification and the burden of pathologists,and improve the accuracy of cancer diagnosis.The main challenge in solving this problem is that,in very large pathological images,the cost of cell-level labeling is huge,and a large number of pathological images only have image-level labels and lack cell-level labels.However,cell-level morphological features is a key factor in tumor diagnosis.Therefore,existing deep learning techniques generally use supervised learning methods to extract full-map representations,rather than cell-level features,for cancer diagnosis.Nevertheless,such methods for extracting full-map representations still have obvious deficiencies:（1）poor interpretability: full-map feature representation cannot explain microscopic features such as tumor cell types,numbers or morphology?（2）high computational complexity: full-map features Extraction needs to fully consider the feature calculation of the entire pathological image.Therefore,this thesis uses image-level labels of regions of interest（ROI）in digital pathological images,which are relatively easy to obtain,and uses deep convolutional neural networks,multi-instance learning,and graph convolutional neural networks to extract cell-level microscopic features in digital pathological images.The expression of environmental tissue structure enables pathological image-level classification,while achieving cell-level positive probability prediction.This work can directly calculate the positive ratio of cells in the image,express the proliferation of cells,and predict the malignant degree of the patient’s tumor.In addition,the characterization of cell-level features in this work can improve the interpretability of tumor-positive judgments,which has important clinical implications.This thesis expands from three aspects below:Firstly,the MR-Cell nucleus detection model was constructed by using the idea of transfer learning.Due to the lack of large-scale pathological image data containing cell annotations.In this thesis,we first train a detection model capable of detecting cell nuclei and apply it to a dataset of pathological images to obtain cell annotations for multi-instance learning.Then,a tumor prediction model for digital pathological images was constructed using deep neural networks and multiple instance learning.Inspired by traditional multi-instance learning,this thesis combines traditional aggregation with neural networks.Extract cell features through neural network and output example classification,use traditional aggregation function to aggregate cell classification,and output the entire pathological image classification.Since the classification of pathological images is derived from the cells in them,the cell positivity rate can be calculated.Finally,a tumor prediction model based on graph convolutional network is constructed.Because the graph convolutional network constructs the features of the graph structure,it calculates the relationship between the graph feature data nodes,and can output the node prediction without the graph node label.Due to the similarity between graph convolution and multi-instance learning,this thesis introduces the method of graph convolution network.The graph network structure that constructs the cell image is expressed by cell features,and the structural features of the graph are learned through the graph convolutional network,so as to predict the classification probability of the whole image.This paper constructs a targeted graph convolutional network and obtains a correct rate of 0.92.