Deep Learning Based Nuclei Segmentation In Pathological Slices Images

Accurate segmentation of nuclei in pathological image analysis is very important.Compared with traditional methods that require hand-crafted feature,CNN can automatically learn the features of nucleus,which is helpful for accelerating the diagnosis of pathological images.However,nucleus segmentation is challenging for the variation of nucleus types and morphology.In addition to the diversity of processing task types and difficulty,one of the biggest challenges in pathological image analysis is the deficiency of medical data.It is difficult to obtain enough and balanced data in medical image processing.In recent years,semantic segmentation and instance segmentation have achieved excellent results in various target tasks such as street view recognition.Inspired by these,this paper mainly studies nuclei segmentation in pathological images with the application of semantic,instance segmentation and self-supervised learning.The corresponding solutions are summarized as follows:1)A multi-scale sematic segmentation framework is designed for nucleus segmentation,which utilizes different receptive field size to extract multi-scale feature.A segmentation accuracy on the Mo Nu Seg 2018 dataset of 80.53% is achieved.2)Self-supervised learning is applied to drive information from the data itself.A framework based on scale-wise triples learning and count ranking is proposed,which allows the network to learn feature representations about the size and quantity of nuclei.3)The instance segmentation method Mask R-CNN is introduced for cell segmentation in the paper.Structure-Preserving Color Normalization(SPCN)and different post-processing methods are used to improve segmentation performance.Competitive result on the MICCAI 2018 CPM cell segmentation challenge is obtained.