Weak Supervised Learning For Cancer Pathology Image Segmentation

Semantic segmentation of cancer pathology images has become an important way to assist doctors in predicting patients' cancer risk and cancer grade.With the outstanding achievements of deep learning technology in computer vision tasks,the fully convolutional networks have gradually become the first choice to solve the problem of medical image segmentation.However,this fully supervised learning approach requires a lot of data and artificially produced pixel-level annotations,which usually takes a lot of time and cost.In order to overcome this shortcoming,this paper proposes a cancer pathology image segmentation method based on weak supervised learning,which only uses image-level annotations to achieve pixel-level prediction,which greatly saves the cumbersome manual labeling.This Thesis first proposes a cancer pathology image segmentation algorithm based on multiple instance learning.Using the method of taking the maximum value,the segmentation score map of the fully convolutional network is converted into a probability value of reacting whether the input image is a cancer image or not,and the cross-entropy loss function is calculated with the image-level label of the input image.With the loss continuously decreasing,the network get trained.The final segmentation probability map of the network is combined with a fixed threshold to obtain the segmentation result of the input image.At the same time,the area of the cancer region segmented by the network is larger than the actual area of the cancer region.This Thesis also introduces an area constraint,which effectively improved the segmentation effect.However,the use of image-level labels as the supervised information is relatively limited in terms of network training,and the gap between the final segmentation results and full supervised learning is relatively large.In order to make full use of image-level labels,based on the significant target area localization ability of convolutional neural networks,this Thesis proposes a cancer pathological image segmentation algorithm based on seed cues.Using the classification network and class activation mapping,it can get the approximate region of the cancer in the pathological image to create a pseudo pixel-level label.At the same time,in order to make the obtained pseudo pixel-level label label closer to the cancer area of the image itself,this Thesis also uses multiple dilated convolution and averaging strategy.We call this pseudo pixel-level label as seed cues.Using generated seed cues and SECNet structures based on multiple loss functions,this Thesis achieve accurate segmentation of the cancer area,greatly reducing the gap between weak supervised learning and fully supervised learning on semantic segmentation tasks.