Research On The Segmentation Of Nuclei In Microscopic Images Based On Feature Alignment Domain Adaption Network

The pathological analysis is the gold standard for the clinical diagnosis of many diseases.The number,morphology,and distribution of nuclei in pathological sections under a microscope can provide the diagnostic and therapeutic basis for clinicians.However,the clinician’s manual judgment based on microscopic images is tedious,timeconsuming,and subject to subjective factors.Therefore,it is very desirable and meaningful to use automatic segmentation methods such as deep learning to achieve nuclear segmentation in microscopic images.However,it is a very challenging task to segment nuclei in microscopic images due to the overlapping and adhesion of nuclei,large differences in shape,uneven color in nuclei,complex background,and insufficient data.Aiming at the above problems,a two-stage nuclear segmentation framework combining domain adaption network and segmentation network is proposed in this paper.In the first stage,the domain adaption network uses the source domain image to expand the target domain image,while the segmentation network in the second stage is used to learn the image features of the amplified target-like domain,and to achieve accurate segmentation of the nuclei in the microscopic image by strengthening the discriminant features of extracted nuclei.Specifically,in the first stage,this paper proposes a Global and Local Feature Alignment Network(named GLFA-Net)to achieve domain adaption nuclear segmentation in microscopic images.GLFA-Net includes a global feature alignment network,local pixel reservation network,perturbation consistency regularization,and global feature alignment loss function.The global feature alignment network is mainly improved based on Cycle-GAN.Domain invariant and domain-specific feature extractors are added to enhance the extraction of specific global features between the source domain and target domain,providing global information guidance for generating the class target domain.The local pixel reservation network strengthens local information in the network by introducing a self-attention mechanism to focus on the local area,and maintains semantic consistency between source domain images and generated target domain images,to effectively distinguish nuclei from complex and chaotic backgrounds regions.In addition,perturbation consistency regularization and global feature alignment loss functions are introduced to dynamically focus on semantic and structural changes and enhance the sensitivity of the discriminator to enhanced data,thus improving the generalization ability and the quality of the generated target domain images.Secondly,the segmentation network in the second stage of this paper is used to segment the expanded class target domain data.Specifically,Mask R-CNN was adopted as the main body in the second stage,and a boundary extraction unit was introduced to make the network pay more attention to the nuclear boundary region.In this paper,the performance of the proposed GLFA-NET is evaluated on the challenging source domain dataset BBBCV039,target domain dataset Kumar,and TNBC,and the results are compared with the most advanced methods.Experimental results show that the proposed method is competitive with the most advanced methods in two target domain datasets.