The Research On Segmentation Of Histopathological Slices Images Based On Deep Learning

Histopathology image is regarded to be the gold standard for clinical cancer diagnosis,the delineation of pathological cancer areas enables pathologists to quickly locate the lesion area for further diagnosis and develop subsequent treatment.But the delineation requires a professional with good knowledge and clinical experience.Due to the large size of histopathology images and the complex cancer features,pathologists have deviations in the delineation of pathological image,then may severely affect their diagnostic consistency and subsequently the ensuing pathological analysis.An automatic and precise WSI tumor segmentation framework to assist pathologists has clinical values for improving diagnostic efficiency and consistent diagnosis accuracy.On the other hand,the automatic segmentation of histopathological slices images requires a large number of pixel-level labels,but many fine annotations of pathological images are difficult to obtain.Starting from the characteristics of the histopathological image,we focus on the network model of multi-resolution attention,multi-scale convolution and semi-supervised learning,and we propose solutions under both fully supervised and semi-supervised conditions.The main work is as follows:(1)We propose a multiresolution attention and multi-scale convolution network(MAMC-Net)for the automatic segmentation of WSI in this paper.MAMC-Net based on U-Net is composed of two modules.We propose a multi-resolution attention module to preserve the details and boundaries of the images by aggregating global contextual information from different-resolution images.Then,three attention modules are respectively incorporated between the encoder and the decoder at each level to focus on features in a specific level related with the segmentation task,which may be useful for identifying the smaller and scattered area of cancer.We design a multi-scale convolution module to capture multi-scale semantic information extracted from the deepest layer of encoder and integrate them with the output feature representation of the decoder to construct multiscale features representation.Finally,we apply the softmax function into the output layer of network architectures for the normalization of the tile-wise logit mapping.Then CRF is exploited to splice the overlapping segmentation maps to preserve the continuities and consistencies in WSI segmentation map.We demonstrate the effectiveness of our MAMC-Net on WSI tumor segmentation tasks using two benchmark datasets.(2)we propose a semi-supervised histopathological image segmentation based on multi-task learning.Cancer region segmentation task and classification task are simultaneously trained using semi-supervised method in this paper.Firstly,we use a limited number of pixel-level labels to train a segmentation network,and then some image-level labels are adopted to simultaneously train the segmentation network and a classification network.The network parameters are optimized by alternating iterative method in the training process of the network.Thus,our method provides a feasible solution to reduce the requirement of label data for deep learning model.Furthermore,we introduce a dynamically weighted cross entropy loss function to train the network,which can automatically allocate the weight of each pixel based on the accuracy of classification prediction.Therefore,the target regions with the low probability of classification prediction are paid more attention in the process of the training,then the details of the cancer regions can be preserved.Experimental results on the breast cancer histopathological image dataset verified that our method outperform other state-of-thearts on the cancer segmentation performance in the case of insufficient pixel-level label data.