Research On Weakly Supervised Segmentation Algorithm For Pathology Image

With the development of deep learning technology,the relevant algorithms based on fully supervised learning have become the main choice in the field of medical images nowadays,and this fully supervised learning method requires a large amount of data as well as pixel-level annotation.Due to the constraints of manual annotation time and cost,large volume of data annotation is usually difficult to obtain in practice.In order to overcome the above-mentioned shortcomings,this paper investigates the problem of pathology image segmentation under weakly supervised conditions based on sparse annotation with breast histopathology images and colon histopathology images as the research objects,and the main research contents are as follows:(1)A pathology image recoloring method based on generative adversarial networks is proposed to address the problem of stained tissue image discrepancy in pathology image datasets.Influenced by the imaging platform,staining platform and stain formulation,digital pathology sections from different institutions differ significantly in color,and this difference greatly affects the performance improvement of the weakly supervised segmentation algorithm.In this paper,the proposed pathology image translation algorithm achieves the control of image chromaticity information in the generation network by introducing color histogram features,and finally generates pathology images with consistent staining "style" to achieve the standardization of pathology image staining.Compared with the traditional staining standardization methods,the proposed method has a significant improvement in the distortion of tissue and texture structure of pathology images.(2)A class activation map optimization method based on spatial consistency constraints is proposed for the problems of rough localization results and misalignment of segmentation accuracy of pathology images by the class activation map CAM(Class Activation Maps)method.Currently,the CAM method is a key technique in weakly supervised pathology image segmentation models for extracting concern maps for target semantic segmentation from classification models.However,the sparse annotation at the image level brings very weak supervised information to the network,so that the attention of CAM methods usually focuses on the high discriminative regions in pathology images and ignores other regions.Therefore,this paper proposes a CAM model optimization method based on spatial consistency constraints,which improves the ability of CAM to finely localize target objects by explicitly introducing additional supervisory information.The experimental results show that the method can steadily improve the quality of the target attention graph and improve the performance of the corresponding weakly supervised segmentation model.(3)A CAM optimization method based on attentional feature enhancement is proposed for the problem of variable shapes and significant differences of glands and lesion regions within pathological images.Since the target attention map is generally obtained from the last layer of the output feature map with the help of the CAM method,there is still room for optimization of the fitting ability to detailed regions such as edge contours.In order to further improve the ability of the weakly supervised segmentation model to finely localize the target object,this paper introduces a self-attentive mechanism to realize the association of features and the construction of long-range dependencies,which in turn improves the learning ability of the model for the contour morphology of pathological images.The experimental results show that the method is effective in improving the segmentation performance of pathology image contour details,further bridging the gap between fully supervised learning and weakly supervised learning in pathology image segmentation tasks.