Research Of Sparse Representation Method For Histopathological Images Classification

Histopathological images classification is a key part of Computer Aided Diagnosis(CAD)systems.Due to the multiple pathological information,complicated dyeing process,and individual differences in patient,the differences of intra-class images may be greater than the differences of inter-class images,which is the main challenge for the histopathological images classification.In recent years,the dictionary learning and sparse representation have received extensive attention from domestic and foreign scholars,and become a key technique for images classification.This paper focuses on the hotspots and difficulties in the classification of histopathological images,conducts detailed research,and has achieved excellent classification results.The main works of this article are as follows:(1)A new classification framework of histopathological images is proposed in this paper,which consists of three parts:a stacked-based discriminative prediction sparse decomposition(SDPSD),a multi-channel joint sparse model and a spatial pyramid match(SPM)model.First,using the SDPSD model,the framework extract the sparse decomposition features of the RGB three-channels in the pathological images,and we clusters them to obtain the dictionary;secondly,the features of RGB three channels are respectively expressed as the sum of two parts:the common components of the three channels and the unique components of each channel,and we obtain sparse representations of various components based above dictionary and establish the multi-channel joint sparse model;finally,combining the spatial pyramid matching,a joint sparse representation coefficients of different levels of image features is obtained,and by using support vector machine to complete the classification task.The validity of the proposed model is verified on the ADL and BreaKHis data sets.(2)Considering that the common component and the unique component are coded over the same dictionary in above classification framework histopathological images,which causes higher similarity of the sparse representation coefficients,and lower discrimination of the joint sparse representation coefficients.So,the mutual Information-based multi-channel joint sparse model(MIMCJSM)is proposed for identifying the types of disease accurately.First,the model extracts the sparse decomposition features of RGB three-channels based on the SDPSD,and using the K-means algorithm,we respectively cluster the features of each channel to obtain the R,G,and B dictionaries;then,exploring mutual information between training samples and three dictionaries,we delete the irrelevant atoms to construct the shared dictionary and three unique dictionaries.Simultaneously,a multi-channel joint sparse model is designed based on a shared dictionary and three sub-dictionaries.Furthermore,in order to represent image feature of different levels,a spatial pyramid matching(SPM)is used into the multi-channel joint sparse coding by applying image spatial information;finally,joint sparse coding coefficients are used to train the SVM classifier for histopathological images classification.The experimental results show that the proposed model has power representation ability and improve greatly the discrimination of coding coefficients.Thus compared with the traditional models,the better classification performance and the power robustness can be obtained.