Nuclear Image Segmentation Based On Attention Mechanism

Pathological examination is regarded as the "gold standard" for cancer diagnosis.The use of computer aided diagnostic systems to assist physicians in pathological examinations can improve the speed and accuracy of diagnosis.In the system,the accurate automatic segmentation of cell nuclei is extremely important for the subsequent calculation of diagnostic indicators and extraction of cell nuclei lesion features.Deep learning-based segmentation methods are currently widely used in nuclear image segmentation tasks.However,in general,deep learning models analyze all the features in the image equally,which makes the learning of the model not targeted.Therefore,it is difficult to improve the segmentation accuracy when facing the problems of the complex image background,low contrast between cell nuclei and background,and overlapping and adhering cell nuclei.While the attention mechanism can reassign weights to the feature map to help the model quickly and accurately acquire key information and ignore irrelevant information to further improve the feature representation of the model.In this paper,we study the attention mechanism and introduce it to the cell nuclei segmentation methods to achieve accurate segmentation.Its main work is as follows:(1)A two-stage cell nuclei segmentation method based on the dual-attention mechanism is proposed to address the problems of low contrast between cell nuclei and background and unclear cell nuclei boundary.The method combines the semantic segmentation network with post-processing operations,which constitute the two stages of this method.In the first stage,Firstly,a dual attention module is introduced in the encoder,which makes the feature extraction focus on the cell nuclei information and ignores the background information.Secondly,the channel interaction module is introduced in the skip connections to provide detailed information about the cell nuclei contours when the decoder recovers the image resolution.Finally,the semantic segmentation results of the cell nuclei are obtained.The cell nuclei are segmented into individual instances by the watershed algorithm in the second stage.This method outperforms the mainstream cell nuclei segmentation methods on both BNS and the self-built ClusterSeg dataset.(2)A cell nuclei segmentation method incorporating attention and multi-task learning is proposed for the problems of complex image backgrounds,overlapping,and adhering cell nuclei.This method first performs feature extraction on the nuclear image,then identifies the cell nuclei instances by the object detection head,and finally segments the cell nuclei based on the object detection results.An effective information-aware module is introduced for feature extraction,which combines global average pooling,attention gate,and attention mechanism to enhance feature extraction of cell nuclei and lay the foundation for the object detection head to be able to accurately identify each cell nucleus instance.An adaptive receptive field segmentation branch consisting of deformable convolution is proposed for segmenting the cell nuclei contours.It forms a multi-task learning framework with the cell nuclei segmentation branch to explore the association features of cell nuclei and contours,and accurately find the cell nuclei boundary,thereby improving the segmentation accuracy.It is experimentally verified that this method outperforms the mainstream cell nuclei segmentation methods on MoNuSeg,CPM17,and the self-built ClusterSeg dataset.(3)A weakly supervised cell nuclei segmentation method based on self-attentiveness is proposed to address the problems of time-consuming and difficult acquisition of pixel-level annotations.The method uses box-level annotations to supervise cell nuclei segmentation to improve model segmentation accuracy in terms of adding additional supervised information to the model and enhancing model segmentation performance.The cross-view similarity is introduced into the siamese network framework as a priori information to add additional supervised information to the model by calculating semantic consistency loss and instance consistency loss.Then the self-attention mechanism is introduced to make cell nuclei instances learn from each other and improve cell nuclei segmentation accuracy.This method outperforms the mainstream weakly supervised cell nucleus segmentation methods on MoNuSeg,CPM17,and self-built ClusterSeg datasets and shortens the gap with fully supervised segmentation methods.In summary,this paper proposes a series of solutions to address the challenges encountered in cell nuclei segmentation in pathological images and verifies the effectiveness of the proposed methods through extensive experiments.The experimental results show that the methods based on attention mechanism in this paper can make the network focus on the features of cell nuclei,thus improving cell nuclei segmentation performance.