Research On HER2 Image Classification Algorithm Based On Convolution Edge Enhancement

Among all kinds of cancers in human,breast cancer has always been one of the most serious diseases that threaten the health of women.However,either human epidermal growth factor receptor 2(HER2)gene amplification or high expression of related protein exists in the patients with breast cancer,which is called HER2 positive breast cancer.As an independent predictor of long-term outcomes,it plays a guiding role in the options of treatment for patients.The HER2 scores is mainly determined by the staining degree and proportion of the invasive cancer cell membrane.In the past,its evaluation based on manual observation.Not only the workload is heavy,but the accuracy depends on the doctor's experience and subjective consciousness largely.Because of that,under the premise of lightening the workload and guaranteeing the precision,this thesis focus on optimizing the models of convolutional edge detector and pathological classification for HER2 image as possible.The main work of this thesis is as follows:(1)Do research on the characteristics and criteria of HER2 pathological images.Through the binarized process in thresholding,a thumbnail is gotten and then manually labeled region will be mixed in order to get the region of interest.On the basis,images are cut in the size of 256×256 and artificial observation is used for labeling in 16000 images without the meaningless images.(2)The combination of traditional feature extraction including LBP,GLCM and HOG,and classical classification algorithm including LP and SVM is adopted.Do the research on using a small amount of labeled data,with different combinations data,to automatically label a large number of unlabeled data,and the experimental results are analyzed theoretically to evaluate the performance of different combined algorithms.Next,preprocessed labeling is done with LBP and SVM,which performed best before.At last,training network needs a lot of labeled data,which is amplified by data augmentation and generative adversarial networks.(3)Aiming at the enhancement of key edge features of HER2 images.By studying SE structure,which can adaptively learn how to distribute the features of weight,a modified model called HED_SE is proposed from HED model.The edge detection is tested in Structured Edge Detection Toolbox to verify the improvement with ODS,OIS and AP.Next,the gray distribution of the dataset is counted so that the appropriate threshold can be selected,which guarantees the most evident differencebetween categories of HER2 images.After covering the key edges of the membrane staining in original HER2 images with HED_SE,a tiny CNN is used for testing the effect of enhanced HER2 images compared with the original images in ROC curve.(4)Studying various special network structures such as dilated convolution,group convolution,Inception,Bottleneck,DepthWise Convolution and Channel Shuffle.Based on these structures,the light-weight and multi-task network is proposed for the classification of scores and magnification of HER2 images with the deep learning library named Keras.Through the experiment,which contains the comparison with several classical neural network in parameters,training time and ROC curve,this thesis proved that the proposed network still has a nearly great effect even with the much fewer parameters.Finally,the local region of interest is analyzed by heat map to test the actual classification effect of the model.