Nuclei Detection Methods For Digital Pathology Images Based On Domain Adaptation

In clinical practice,histopathology images are the gold standard for diagnosing most diseases.As an important part of histopathology image analysis,automated nuclei detection can assist pathologists in disease diagnosis,thus improving the efficiency and accuracy of pathology diagnosis,which has important research significance and wide application prospects.At present,nuclei detection methods have achieved some success in histopathology images.However,these methods require that the training and testing data are obtained from the same probability distribution,otherwise the performance of the model will suffer from severe degradation,which is called the domain shift problem.In recent years,the rapid development of deep learning based domain adaptive technology provides new opportunities to solve the above problems.However,the existing cross-domain nuclei detection methods do not consider the distribution differences of pathological images in the feature space,resulting in limited generalization ability of detection models.In addition,due to the small size of nuclei,nuclei instance features lack the discrimination ability,which affects the improvement of cross-domain nuclei detection performance.Finally,some nuclei in histopathology images exhibit deformation,blurring,etc.,resulting in deviation in the overall distribution of instance features and reducing robustness in aligning nuclei related features.Therefore,this paper closely focuses on the domain shift problem of the nuclei detection task in histopathology images,and a variety of novel domain adaptive methods are proposed on this basis.The main contributions of this study are as follows:(1)Aiming at the problem that existing cross-domain nuclei detection methods ignore the difference in feature distribution of images,this paper proposes an instance feature alignment based cross-domain nuclei detection method IFA.Firstly,a global and local bilayer-level feature adversarial alignment framework is designed,which utilizes the local alignment component to align the instance features of nuclei regions while aligning the global features of the whole images,so as to effectively reduce the difference in feature distribution of the histopathology images.In addition,a temporal ensembling based nuclei localization module is proposed to accurately extract instance features in target domain images,thereby better assisting the instance feature alignment process.IFA was evaluated by using multiple performance evaluation metrics,which shows that the method has favorable cross-domain nuclei detection capabilities.(2)In order to enhance the discriminating ability of nuclei instance features in histopathology images,on the basis of previous research,this paper proposes a selfattention based cross-domain nucleus detection method GLAFA.Specifically,a location-aware self-attention module LocSA is designed to refine instance features by fusing all instance features of nuclei regions in the histopathology,which makes instance feature contain detailed information from other nuclei regions to effectively enhance its discrimination ability and achieve better local alignment effect.Moreover,this paper introduces pyramid pooling technology into the self-attention module,which alleviates the problem of high computational complexity by reducing the instance feature dimension.Experimental results in multiple domain adaptive scenarios show that the GLAFA method can further improve the performance of cross-domain nuclei detection.(3)Aiming at the problem that self-attention based feature refinement affects the specific expression of instance features,a nuclei graph based feature refinement and selection alignment method called GNFA is further proposed in this paper.This method introduces a nuclei graph convolutional network,which avoids the interference with nuclei information unrelated to the central node by fusing the instance features of spatially adjacent nuclei regions,thus ensuring the specific expression of local instance features and enhance the actual effect of feature alignment.In addition,in order to solve the problem that some nuclei with poor quality in histopathology images affects the alignment of node features,the importance learning module is utilized to select highquality node features for adversarial alignment,so as to effectively improve the robustness of feature alignment.In this paper,GNFA is compared with existing domain adaptive methods for experimental and performance evaluation,which fully verifies the superiority of this method in cross-domain nuclei detection tasks.