Classification Of Pathological Images Based On Optical Scanning

At present,the modernization process in the field of pathology in China is very slow,and the efficiency of traditional working methods is low.The doctors in pathology department are difficult to train,resulting in a general shortage of pathologists.In order to improve the status quo of Pathology,this paper uses convolutional neural network algorithm to design a set of pathological diagnosis system to assist the pathologist diagnosis,to speed up the digital automation of pathology.Using an optical scanner to digitize the pathological sections and then using the algorithm of convolutional neural network to automatically detect them is a very promising scheme at present,and also a goal that can be realized and possibly be used as a clinical assistant.Some image recognition algorithms based on convolutional neural network imitate human vision,extract relevant features in digital pathological images,and make algorithm judgment on images.In this paper,the digital pathological images of breast sentinel lymph nodes are taken as the research object,and the classification and segmentation methods of digital pathological images by convolutional neural network are deeply analyzed and designed.Through the comparison of many experiments and the analysis of specific problems,the classification algorithm of patch stage with resnet-50 as the backbone network and the inference system of the whole digital pathological image using FCN idea are constructed.The main contents of this paper are as follows:(1)Based on the Convolutional Neural Network(CNN)of intraoperative digital pathological images of frozen lymph nodes and the relevant algorithm of neural network,this paper designed a very effective two-stage algorithm for segmentation and classification of ultra-high resolution digital pathological images.The first stage is to extract the region of interest from the oversize image and cut it into small images of a certain size for classification.In the second stage,the results of the first stage were spliced into a heat map proportional to the size of the original image in a certain order,and the whole pathological image was classified using the heat map.(2)Use the idea of full convolutional network(FCN)for reference to accelerate the process of large resolution image inference.Mainly by replacing the full connection layer at the end of the model in the first stage with the convolutional layer,the calculation amount of the whole pathological image inference can be reduced by means of the convolutional sliding window,so as to reduce the inference time.(3)In this paper,a domain adaptive algorithm based on the Generative Adversarial Networks(GAN)is designed by studying the difference and connection between the pathologic image classification project of frozen lymph nodes during operation and the pathologic image classification project of lymph nodes after operation.This algorithm successfully integrated the intraoperative freezing model and the postoperative model,and achieved excellent results.