【作者】 孙树娇； 郑钰山； 姜志国； 谢凤英；

【Author】 SUN Shujiao;ZHENG Yushan;JIANG Zhiguo;XIE Fengying;Image Processing Center,School of Astronautics,Beihang University;Beijing Advanced Innovation Center for Biomedical Engineering,Beihang University;

【机构】 北京航空航天大学宇航学院图像中心； 北京航空航天大学医工交叉创新研究院生物医学高精尖创新中心；

【摘要】 组织病理切片检查是癌症诊断的金标准。然而,癌症组织结构变化各异,诊断耗时长且易于出错,不同病理专家对同一切片的诊断结果存在多样性,为癌症的精确诊断造成了困难,研究应用于病理全切片的人工智能辅助诊断方法具有长远意义。近年来,基于全自动显微镜的组织切片成像技术在医学病理领域得到快速普及,使基于病理切片图像分析的计算机辅助诊断方法成为该领域的研究热点。基于深度学习的病理全切片图像分割是该领域的主要研究方向之一,利用深度学习的方法对病理全切片图像进行分割,可以对癌症区域进行准确定位,为后续癌症的诊断提供辅助信息。然而,病理全切片图像分辨率高,尺度大等特点为癌变区域自动分割任务带来了巨大的挑战。现有分割方法主要利用高放大倍数下的全切片图像进行建模,在测试的过程中耗费较大的计算量,致使自动处理过程需要较长的时间。针对这些问题,本文提出了一种基于级联卷积神经网络的全切片图像快速分割方法。本文方法涉及3个神经网络之间的连接和训练,每个网络均为U-Net网络结构,分别用于分割放大倍数为10倍、20倍和40倍下的切片。分割过程中,切片首先通过低倍率网络,并根据低倍率网络的输出判断各区域的后续处理方式,其中具有高置信度的阳性和阴性区域直接输出结果,不再进行后续计算;预测置信度较低的部分送往高倍率网络进行进一步分割。另一方面,级联网络训练中针对不同类别样本采用不同的监督策略,提出了一种癌变组织敏感损失函数(Cancer Sensitive Loss,CSL),促使低分辨率网络以较高的置信度滤除非癌变组织区域,进一步平衡了级联网络的分割精度与速度。本文方法在ACDC-LungHP竞赛数据集上进行了验证,实验结果及对比表明,本文所提出的级联策略,在大幅减少了计算量的同时,一定程度上提高了分割的精确度。本文的主要贡献总结如下:①提出了一种针对病理全切片图像分割的级联策略;②提出了一种癌变组织敏感损失函数(Cancer Sensitive Loss,CSL)。本文提出的级联策略理论上可用于任何图像分割网络以处理具有高分辨率、大尺度等特点的图像。更多还原

【Abstract】 Cancer diagnosis still relies on histopathology.However,the manual analysis of histopathological whole slide images(WSIs) is a time-consuming task for pathologists and often suffers from errors and intra-observer variability because of the diversity of cancerous organization.It’s significant to develop automatic analysis of histopathological WSIs based on artificial intelligence.With the development of digital pathology,histological sections can be scanned rapidly using advanced micro-scanners and stored as digital WSIs.Due to the accessibility of large amounts of WSIs,the computer-aided diagnosis methods based on histopathological WSIs have become popular,especially the methods for cancerous regions segmentation based on deep learning technology.The segmentation results can accurately localize the cancerous regions and provide auxiliary information for the following diagnosis.Automatic segmentation of histopathological WSIs is challenging due to the characters of WSIs,high resolution and large scale.Most existing segmentation algorithms,to achieve high accuracy,process images at high magnification and consume excessive calculation.To address these problems,in this paper,we propose a cascade strategy for fast segmentation of WSIs based on convolutional neural networks.Our segmentation framework based on cascade strategy consists of 3 associated networks.Specifically,3 U-Net structures trained with different samples with increasing magnifications,10×,20× and 40×.During the segmentation processing,the WSI is first fed to the network trained with low magnification samples.Then,the positive and negative regions with high confidence are output directly and,on the contrary,the remaining regions are fed to the next network trained with higher magnification samples for further processing.Simultaneously,we propose a cancer sensitive loss(CSL) function which can help the network trained with low resolution samples to filter non-cancerous regions with high confidence and meanwhile to strike a balance between segmentation accuracy and speed.We designed the experiments based on dataset provided by ACDC-LungHP challenge and the results demonstrate that segmentation network based on cascade strategy consumes less calculation even improves the segmentation accuracy.The main contributions in this paper includes:①A cascade strategy is proposed for the segmentation of histopathological WSIs.②We propose a cancer sensitive loss function for the imbalance of positive and negative samples.The cascade strategy is still effective for other segmentation networks when processing images with high resolution and large scale.更多还原