Nucleus Segmentation Based On Deep Learning

Automatic accurate segmentation of cell nuclei in pathological images plays an important role in cancer diagnosis.Aiming at the problem of low accurate segmentation results caused by large differences in the size of nuclei in pathological images,a new size adaptive automatic segmentation method based on Mask R-CNN is proposed.The method is divided into three steps.First of all,the pathological image is pre-segmented,and then the pathological image is sorted out into a predefined class according to the pre-segmentation results.Finally,the pathological image is segmented again using the model of the corresponding class,and the final nucleus segmentation result is obtained.Among them,the model of the corresponding class is trained by using the image of the classified training set.Classification is based on clustering,which is based on the average size of nuclei in the image.The experimental data of this paper is from the 2018 data science bowl(2018DSB)challenge dataset,which includes one training dataset and two test datasets.Compared with Mask R-CNN,U-Net and the latest nuclear segmentation methods,the experimental results show that this method is superior to these methods.The DSB score are 0.471 and 0.580 respectively,and the average of F1 score are 0.780 and0.864 respectively.Although the result is lower than the best method based on DSB challenge dataset,the method proposed in this paper is easier to implement,and the training data only comes from DSB original training dataset.In order to verify the effectiveness of the proposed method on the new dataset,we also compared with Mask R-CNN and manual counting result on the Nuclei Counting100 dataset.The experimental results show that proposed method is greater than Mask R-CNN with an average of 108 differences.Although the results of the previous study have achieved good and effective results,but the segmentation effect is poor for nuclei with completely overlapping or large overlapping area.Aiming at this problem,we first proposed a multi-task model based on the centre area of the nucleus and contour perception,but the counting results of this model are not good.Therefore,in order to improve the above method,a new method of counting nuclei based on classification is proposed.This method is to train a new nucleus center detection model on the basis of previous study.In this paper,we first use the euclidean distance transformation method to transform the segmentation results,and then use the center area detection model of the corresponding class to get the center area in the distance transformation graph,so as to achieve the counting of overlapping nuclei.The center area of nucleus used in the training model is obtained from the distance transformation map of the training image through a fixed threshold.Compared with the method proposed in the previous study,Mask R-CNN and the latest method,the experimental results show that the average difference between the counting results obtained by the proposed method and the real results is the smallest,which is 15.26 and 6.99,respectively.The comparison of this method with Mask R-CNN and the method proposed in the previous study on the Nuclei Counting100 dataset shows that the mean difference of this method is 119.22,which is better than Mask R-CNN,but lower than the method of the previous study.The R-square value of this method is 0.6499,which is similar to the method in the previous study,but is significantly better than Mask R-CNN.The two methods proposed in this paper have good results in nuclear segmentation and counting,especially for images with large differences in nuclear size distribution.At the same time,this method has a good reference for the detection of large size difference target or large shape difference target.