Ultra-high Resolution Pathological Image Segmentation Algorithm Based On Weakly Supervised Learning

Histopathological images are important for doctors to locate the tumor area and diagnose cancer.With the development of deep learning,automated pathological image analysis methods have been widely used.However,the resolution of histopathological images is very large,and it is very difficult for pathologists to accurately locate the cancerous area.Therefore,combined with deep learning technology to automatically locate the location of the cancerous area can greatly reduce the workload of the doctor.Since the fully-supervised semantic segmentation method requires a large number of pixel-level labels,and the acquisition of pixel-level labels requires a lot of manpower and material resources,the use of weakly-supervised segmentation methods can greatly reduce the cost of label acquisition and reduce the cost of manually making label.This paper firstly proposes a fully-supervised cascaded deep-supervised pyramid pooling network to improve the segmentation effect of ultra-high-resolution pathologi-cal images.Firstly,the downsampling method is used to extract the context information of the image,so that the network can accurately judge the benign and malignant of pathological tissue,and generate rough segmentation results.Then,the cutting method is used to input the high-resolution image and its context information into the network to realize the accurate segmentation of the boundary contour of the cancerous region.In the network training process,add an auxiliary convolution layer to the middle feature layer,calculate the loss with the ground truth,optimize the middle feature layer,and modify the pyramid scene parsing network,using the dilated convolution in the feature extraction module to expand the receptive field and improve the effect of segmentation.In view of the high cost of pixel-level labeling,this paper proposes a pseudo-label correction network based on weakly supervised learning to perform two-norm regular-ization on the class activation maps generated before and after data enhancement.This operation can generate a more robust class activation map,reduce the gap between the classification network and the segmentation network,and make the classification net-work in the feature extraction part and the feature map generated by the segmentation network converge.This article compares the method proposed in this article with other methods on the colorectal pathology dataset Digest Path 2019,and proves the effectiveness of the two segmentation methods proposed in this article from qualitative and quantitative analy-sis.The cascaded deep network trained in a fully-supervised manner can improve the segmentation effect and optimize the border and contour of the cancerous area.The pseudo-label correction network trained by the weak supervision method can accurately locate the cancerous area and generate a more reasonable saliency area when only use image-level labels.In this paper,a segmentation system of histopathological images is implemented,and the segmentation algorithm proposed in this article into it.In addition to this,a va-riety of traditional image processing methods that are helpful for pathological diagnosis are implemented and integrate into the system.This system promotes the research to product can improve the diagnosis efficiency of pathologists.