Research For Image Detection And Classification Algorithm Based On HE Staining Slides

It's essential to have a treatment early because of the harm of cancer,which is self-evident.Computer-aided diagnosis of pathological images can reduce the workload of pathologists and help to decrease misdiagnosis rate,which has important clinical value.The classification and detection of pathological images are two basic tasks in digital histopathological image analysis.It's pivotal to detect the lesion area on the Whole Slide Image(WSI)accurately.Histopathological images with Hematoxylin-Eosin(HE)staining,whose tissue structure and cell morphology are easy to observe.Another advantage is that HE staining slides can be stored for a long time.Based on this,the aim of this study is to study the image detection and classification algorithm based on HE staining slides,which includes the training of classification models and the detection of pathological images.Two methods based on deep learning are adopted to train the classification models which are evaluated to select the better for pathological image detection.The detection of pathological images involves two objects,one is large-scale image,and another is WSI.In the detection for WSI,this paper improves the original detection algorithm for its inadequacies.Then a large number of experiments are done to prove the superiority of the improved algorithm.The main research of this paper is as follows:(1)Based on the characteristics of pathological images with HE staining,the region of interest was extracted from WSI according to the color difference.Under three kinds of magnifications,three kinds of sample sets were prepared according to the pathologist's annotation.Color normalization of images was done due to color differences among slides which have different sources.(2)This thesis adopted two ways to train classification models.First,based on the study of DenseNet and SE-DenseNet structure,the lightweight and memory-optimized DenseNet network was used to train Convolutional Neural Networks(CNN)models.Second,based on transfer learning to train classification models,9 common CNN models such as InceptionV3 were used to extract features of images.The Next step was to train a SVM,RF classifier or train a fully connected layer directly to generate classification model.For each way to train models,three classification models were trained under three kinds of magnifications.The classification model with best performance,which was selected by evaluating the performance of all models,was used for pathological images detection.(3)This thesis studied the detection algorithm of large-scale pathological image and WSI.First,for the detection of large-scale pathological image,combine classification models from multiple magnifications for analysis.Based on the detection results under low magnification,detect the area with low confidence under higher magnification.Then,SLIC algorithm was used to smooth the edge of the detected cancerous area.Compare the detection results with the pathologist's annotations and evaluate the results quantitatively.Second,for the adaptive sampling detection for WSI involves the following steps: tiles were extracted from the WSI using sampling algorithm;the classifier was applied to each tile;the prediction produced by the classifier was used to build an interpolated probability map;select the positions of next sampling points according to the gradient map of probability map.This paper improved the original algorithm for its inadequacies,a three-stage sampling method was proposed and the iterative stopping condition based on sampling density was set to improve the efficiency and precision.A large number of experiments were carried out to evaluate the performance of the improved algorithm.