Histopathological Images Classification Based On Discriminative Dictionary Learning

Histopathological image contains abundant pathological information of organisms,which have immeasurable value for the development of medicine research.Therefore,histopathological images classification has attracted great attention in academic area.In recent years,many achievements have proved the importance of discriminative dictionary in image classification.This paper focuses on how to learn a discriminative dictionary,and further achieve better classification performance of histopathological images.The main work of this paper are as follows:(1)Aiming at the problem that there is a high correlation among sample features in histopathological image classification,a discriminative dictionary learning algorithm with low-rank constraint(LRCDDL)is proposed.The traditional algorithms generally only consider the low-rank of sparse coding.LRCDDL not only optimizes the reconstruction performance of sub-dictionary for the same and different class training samples,but also exploit low-rank constraints on the class-specific subdictionary.This strategy can reduce the correlation among the class-specific subdictionary atoms and promote the independence of atoms,which helps to learn a discriminative and compact dictionary.The experimental results on ADL and BreaKHis dataset show that LRCDDL can achieve higher classification accuracy than the existing algorithms.(2)Although the above LRCDDL algorithm has achieved good results in the histopathological images classification,the classification results of LRCDDL algorithm are not satisfactory when classifying noisy histopathological images.In order to address this problem,a discriminative dictionary learning algorithm with pairwise local constraints(PLCDDL)is proposed.Unlike most algorithms that directly use training samples to construct Laplacian matrices,PLCDDL algorithm trains a pair of adaptive Laplacian matrixes by using the class-specific dictionary.On the one hand,this strategy can reduce the influence of abnormal data and noise on image classification results in histopathological images,which helps enhance the robustness of the learned dictionary.On the other hand,it can update the Laplace matrix adaptively with the update of the dictionary.In addition,pairwise local constraints constructed by Laplacian matrixes can enhance the similarity of the coding coefficients in the same class but suppress the ones in different classes,which can further improve the discriminability of learned dictionary.Therefore,PLCDDL is helpful to learn a high-quality discriminative dictionary.Experiments on ADL and BreaKHis datasets demonstrate the superior classification performance of the proposed PLCDDL algorithm.