Multi-Tissue Nuclei Segmentation From Pathological Images Based On Dual U-Net

Cancer is the second leading cause of death in the world,and the incidence is increasing year by year.Pathological diagnosis is the "gold standard" for cancer diagnosis.The size,shape and distribution of nuclei in pathological images are highly correlated with cancer diagnosis,grading and prognosis.However,manual labeling of nuclei in pathological images has problems such as heavy workload,poor reproducibility and high experience threshold.Therefore,there is an urgent need for developing an accurate automatic nuclei segmentating method.Up to now,due to the scarcity of labeled pathological images,uneven color distribution of different pathological images and overlapping distribution of nuclei,the segmentation of pathological images remains a huge challenge.Based on the existing segmentation methods,this paper proposes an innovative method for nuclei segmentation in response to some problems existing in nucleus segmentation in pathological images.In this article,based on the multi-task learning method,we designed a double U-Net network to perform nucleus segmentation.Aiming at detecting the nucleus boundary,one UNet in the dual U-Net network is used to predict the nucleus area,and the other U-Net is used to predict the nucleus boundary area.There is a feature fusion and interaction block between the two U-Nets to exchange information between tasks.At the same time,in order to make the network pay more attention to the cell nucleus boundary area,we introduce an attention mechanism in the feature fusion and interaction block of the decoding stage,so that the nucleus segmentation task can make full use of the information in the nucleus boundary detection task.In view of the scarcity of nuclear annotation data in pathological images,we adopted the method of model-agnostic meta-learning to learn features with universal applicability,thereby improving the generalization ability of the model.The experimental results show that our model has comparable performances with state-of-the-art models.And when the training data drops to 80%,our model performs better,indicating that our model is more suitable for situations with fewer training samples.At the same time,in domain adaptation and domain generalization experiments,our model outperformed existing models,which proved that our model has a stronger generalization ability.The ablation experiment also proved the effectiveness of our proposed multi-task meta-learning method and the information fusion module based on the attention mechanism.At the same time,in the extended experiment,we also explored the influence of self-supervised methods and multi-scale input on the performance of the model.