Study On Mitosis Recognition In Breast Histopathology Images

In breast cancer pathology images,mitosis count is one of the important indicators to evaluate the breast cancer grading.The existing medical methods consume much time and have high professional requirements which detect it manually.Therefore,there are more and more computer-aided diagnosis methods for mitosis recognition.For breast cancer cell have distinct shapes and complex mitotic processes.At present,researches on mitotic identification methods by domestic and foreign mainly focus on feature extraction and classification.These processes take a lot of time and the recognition accuracy is not high.In this paper,three different detection and classification methods are designed and implemented.The main research contents of this paper are as follows:1)Aiming at traditional mitotic detection methods deal directly with RGB images,ignoring some useful features,a new computer-assisted mitosis detection algorithm is proposed based on multi-color channel feature fusion(MCCFF).According to six color channels features are extracted,and multi-channel features obtained are serial fused.Firstly,the cell nucleus is segmented into the candidate set in the pathological image of breast cancer.Then,the features are extracted and fused on different color channels,respectively.Finally,the improved minimum distance classifier is used to obtain the detection result.The experiment takes the data of ICPR2012 mitosis detection competition,and the detection precision reaches 0.823.The results show that the proposed algorithm(MCCFF)is superior to other scholars' algorithms.2)Aiming at the traditional mitotic classification methods the feature extraction process needs to extract a large number of features,and the calculation is complex and time-consuming,this paper presents a computer-aided mitosis detection algorithm based on convolutional neural network combined with ZCA whitening.Firstly,mitotic nuclei and non-mitotic nuclei are segmented as a candidate set.Secondly,ZCA whitening method is used to process the images.Then convolution neural network(CNN)is employed to extract features of the image data layer by layer.Finally,softmax is used to process the network output feature vectors for classification.The effectiveness of the ZCA-CNN algorithm is analyzed by the experimental results.Compared with LSVM and traditional CNN algorithm,the classification accuracy is 0.9115.3)Aiming at the problems of long training time and limited calculation in traditional network model,this paper presents a computer-aided mitotic classification algorithm based on improved depth convolutional neural network(GAA-CNN).First,the candidate set is segmented.Then the deep features of the candidate set are extracted using the improved depth convolutional neural network,and GPU acceleration is used.Finally,the output feature vector is processed by Softmax to obtain the classification result.Through the comparison of the experimental results,the accuracy rate of this method on the breast pathology cell image data set reaches 0.9544,which has a better classification model effect compared with the ZCA-CNN algorithm.In this paper,three new mitotic detection and classification identification methods are proposed,based on the existing problems of the traditional mitotic identification methods.The comparison between the proposed algorithm and traditional algorithm shows that the proposed algorithm has good experimental results.This mitotic recognition computer-aided diagnosis method can be applied to the field of clinical medicine,and provide a lot of help for the diagnosis of early breast cancer.