Nuclear Segmentation Of Pathological Image Based On Deep Learning

The nuclear segmentation of pathological image is of great significance for cancer diagnosis,rating and prognosis.Although the nuclear segmentation of pathological image based on fully supervisied method of deep neural network has made remarkable achievements,the segmentation of pathological image based on fully supervisied method of neural network requires pixel level annotation of pathological image,which requires huge amount of annotations and the cost of annotations is too high.Because of the differences in the color and morphology of the nuclei of different cancer pathological images of different tissues,different disease and different tissue pathological image segmentation is also a great challenge.In this paper,we propose a cell nuclear segmentation framework based on weaklysuperviced learning for different cancer pathological images.Firstly,six methods are used to generate pseudo labels,including levelset,grabcut,watershed method,geocut,circle drawing and bounding box drawing.Specifically,these six methods are used to segment single cells in the bounding box,and the segmentation results are put back to the corresponding positions of the original image.The results show that level set,circle drawing and bounding box drawing can achieve good results.Compared with the ground truth,level set achieves the precision of dice = 0.8790.Secondly,we compare the results of nuclear segmentation of pathological images of different tissues by FCN,deeplabv3 +,UNet and other network structures,and propose our own segmentation network DB-UNet.The results show that our network basically achieves the best results of the fully supervised segmentation with less parameters.After that,we use DB UNET and the generated pseudo labels to segment the nuclei of cancer pathological images of different tissues under weak supervision.The results show that the level set method and geocut method achieve the best results.Finally,we choose the best three networks and the best pseudo lables,and use assembling method to iteratively fuse the pseudo labels.Finally,we use the fusion pseudo labels training to get the DB-Unet proposed in this paper,and the results basically achieves the same precision as the DB-Unet network trained by ground truth.Conclusion: in this paper,we propose a cell nuclear segmentation framework based on weakly supervisied learning for different cancer pathological images.The cell nuclear segmentation has basically reached the results of fully supervisied segmentation,which greatly reduces the cost of data annotations,and also provides a new idea for cell nuclear segmentation.