Classification And Segmentation Of Breast Cancer Pathological Images Based On Deep Learning

In clinical medicine,pathological images analysis is the gold standard for cancer diagnosis while it is hard to analyze them.Diagnosis result was given by pathologists based on pathological images,which is time-consuming,labor-intensive,inefficient and may cause misdiagnosis.Currently,deep learning is emerging in the field of image recognition,which provides a new idea for analyzing pathological images.In this paper,deep learning is used to classify tumor types and segmentation of cancer region in breast cancer pathology images.C omputer-assisted diagnosis reduce the heavy work of pathologists and let the probability of misdiagnosis down.The main work of the thesis is as follows:1.Classification of dataset BreakHis breast cancer pathology images based on improved VGG16 network.For the convolutional neural network training,a large number of samples are needed but the dataset BreakHis is a shortage of pathological image.The common image enhancement technology and the improved image enhancement technique based on histogram equalization are used to process the pathological image so that the dataset can be augmented.The VGG16 network based on migration learning is used to classify breast cancer pathology images avoiding the redesign of network model and retraining the network model.A series of experiments show that the deeper the fully connected layer of the VGG16 network,the higher the classification accuracy.However,it is difficult to improve the classification performance of VGG16 network by adding a full connection layer.A scheme of combining two full connection layer features is proposed to improve the VGG16 network and the best combination of two fully connected layers is explored through a series of experiments.Finally,the experiment shows that the improved VGG16 network can improve classification accuracy.2.Classification accuracy was improved by residual network guided by clustering algorithm.In this section,the clustering algorithm is introduced before the residual network and then the residual network guided by is designed.The pixel of breast cancer pathological images is clustered by the clustering algorithm in the network which can mine the hidden information.Meanwhile,the residual network as a basic network guided by clustering algorithm is trained to solve deep network degradation problems in addition to adding batch normalization to speed up training.Finally,the experiment shows that the residual network can improve the classification accuracy to some extent compared with the residual network.Compared with the improved VGG16 network,the residual network guided by clustering algorithm can improve the accuracy of the image classification under certain magnification.3.Segmentation on the pathological image of the dataset Camelyon 16 based on the segmentation network.Considering that the whole side image contains a large amount of irrelevant background,the Otsu algorithm is firstly used to segment the region of interest.Due to the complexity of the two-dimensional Otsu algorithm,a fast calculation method for the Otsu algorithm is given by the thesis.Because the region of interest of whole side image is purple or red and the background region is other colors which inspires us to introduce color element in the Otsu algorithm and then the Otsu algorithm for segmenting the region of interest of the pathological image is proposed.Finally,the segmentation network is designed based on the classification network.It is difficult to train the whole side image by the network so that the samples and labels are obtained by cropping the whole side image to train network.Finally,the segmentation result of the whole side image is obtained by splicing the image patch segmentation result having been achieved that cancer region is accurately located.