Nuclei Segmentation Based On Self-Supervised Learning Method

Pathological examination is the gold standard for disease diagnosis.The nuclei are important considerations in pathological examination,such as the shape,size and number of nuclei.Therefore,automatic nuclei segmentation is of great value in clinical auxiliary diagnosis and medical research.However,in the field of automatic nuclei segmentation,due to insufficient labeled data,it is difficult for the model to fit the invariance among the samples in the real world.Therefore,it is urgent and meaningful to study how to use a large amount of unlabeled data to cover the invariance.In addition,the feature expression ability of the model is also significant,which is directly related to the performance of the model on tasks.In recent years,the self-attention mechanism has been widely used in the field of computer vision to improve the feature expression ability of the models.However,the self-attention mechanism is limited by its excessive complexity,which requires a trade-off between cost and range when capturing long-range features.Aiming at the problem of insufficient labeled data,this paper proposed a data augmentation method and a self-supervised learning method based on the characteristics of histopathological images and nuclei segmentation,which can better utilize unlabeled data to assist the downstream nuclei segmentation task.Aiming at the problem that convolutional neural networks cannot capture long-range features,and the self-attention mechanism is limited by its complexity,this paper proposed a framework that can capture long-range features based on the characteristics of semantic segmentation model.The work of this paper includes the following three aspects:(1)A data augmentation method is proposed.This method takes into account the characteristics of both histopathological images and segmentation task,which fits well with nuclei segmentation.By using this method to construct positive samples in contrastive learning and constraining the model with contrastive loss,the model can learn the invariance between positive samples.This heuristic constraint enables the model to be trained without labels,thus enabling the use of large amounts of unlabeled data.By comparing with five models and six methods of the same type,this method outperforms other models and methods of the same type in all metrics,such as m Io U,in all three test sets.(2)A self-supervised learning method based on style transfer is proposed.This method assumes that the variability of image styles mainly originates from the variability of stain colors,and model learns the features of each stain through style transfer,thus facilitating the downstream nuclei segmentation.The source domain of transfer is the color normalized image and the target domain is the raw image,i.e.,the raw image is used as the label,thus enabling the use of the large amount of unlabeled data.Compared with the pre-training method based on contrastive learning,this method pre-trains the entire segmentation model and enables the input distributions of the pre-training and fine-tuning consistent.This method was compared with five models and other four self-supervised learning methods,and finally outperformed the other models and self-supervised learning methods in all metrics such as m Io U in all three test sets.(3)A model based on the mean feature vector that can capture long-range feature is proposed.The model uses the result of image segmentation to recapture the relationship between the features in encoder.Thus,the model does not require feature fusion by calculating the similarity between each pair of feature vectors as in the case of the self-attention mechanism,which therefore reduces the computational overhead.In the experiments,the model was trained,validated and tested on three datasets,and data types include histopathological images,retinal images and CT images of brain.The model was compared with several models on the above three datasets and the final experimental results showed the superiority of this model.