Research On Nuclei Segmentation Of Pathological Images Based On Rotation Equivariant And Multi-level Feature Aggregation Neural Network

The histopathological analysis is the gold standard for assessing the presence and many complex diseases in clinic,and the knowledge of professional pathology is required for the pre-prognosis and diagnosis of diseases.As one of the essential part of tumors,the shape,staining,and tissue distribution of the nuclei plays an important role in tumor diagnosis.Due to the heavy reading task and the subjectivity of doctors,the research of automatic nuclei segmentation task is the focus of scholars and experts in recent years.However,due to nuclei congestion and possible occlusion,automatic nuclei segmentation task remains challenging.In recent years,we have paid attention to the cutting-edge technology of automatic image segmentation,deep learning,whose performance in some problems has surpassed the effect of human eye recognition,and it has been implemented in various applications in daily life.In order to overcome the above difficulties,we adopt a deep learning method to study the nuclei segmentation of pathological images.Contributions to this paper are as follows.(1)Considering the inherent rotation equivariant of digital pathological images and the semantic gap between shallow and deep features in the segmentation model based on encoder decoder structure,this paper proposes a core segmentation method based on rotation equivariant multi-level feature aggregation neural network(REMFA-NET).We introduce rotaion equivariant convolutional neural networks to reduce the complexity of samples and increase the expression ability of the network by using the rotational symmetry.(2)To eliminate the semantic gap between shallow features and deep features in encoder-decoder structural models,we propose a multi-level feature aggregation strategy based on U-Net 3+.It includes three improvements: decoder,long skip-connection and semantic enhancement module.Specifically,(1)the up-sampling operation of magnifying low resolution features will cause a sudden oscillation,which makes the model training unstab,and some important information will be lost in the process of magnifying low resolution images.Therefore,we design a new decoder module,which fuses low-frequency features and high-frequency features through residual learning to recover pixel level prediction more accurately;(2)in the ordinary U-Net structure,the effective local field of view is limited by the number of convolutions in the encoder,so the accurate boundary features can not be extracted.We propose an improved long skip-connection method,which adds continuously superimposed residual blocks in the long jump connection to make the decoding stage obtain more rich underlying features and larger receptive field;(3)In feature extraction,continuous down-sampling leads to the loss of information,which affects the image quality recovered by the decoder.According to the shallow features of corresponding edge,we design a feature enhancement block to aggregate the shallow semantic information to enhance the shallow features and ensure the robustness of feature fusion.(3)We evaluated our REMFA-NET on the challenging Kumar dataset and the Mo Nu Seg dataset released in the 2018 MICCAI challenge and compared the results with the state-of-the-art methods.The experimental results show that our method has strong competitiveness in the nuclei segmentation task of histopathological images.