基于深度学习的数字病理图像分割综述与展望

数字病理图像分析对于乳腺癌、前列腺癌等良恶性分级诊断具有重要意义，其中组织基元的形态和目标测量是量化分析的重要依据.然而，由于病理数据多样性和复杂性等新特点，其分割任务面临着特征提取困难、实例分割困难等挑战.人工智能辅助病理量化分析，将复杂病理数据转化为可挖掘的图像特征，使得自动提取组织基元的定量化信息成为可能.特别是随着计算机计算能力的快速发展，深度学习技术凭借其强大的特征学习、设计灵活等特性在数字病理量化分析领域取得了突破性成果.本文系统概述目前代表性深度学习方法，包括卷积神经网络、全卷积网络、编码器—解码器模型、循环神经网络、生成对抗网络等方法体系，总结深度学习在病理图像分割等任务中的建模机理和应用，并梳理了现有方法的方法理论、关键技术、优缺点和性能分析.最后，本文讨论了未来数字病理图像分割深度学习建模的开放性挑战和新趋势.

Overview and Prospect of Deep Learning for Image Segmentation in Digital Pathology

The quantitative analysis of digital pathology images plays a significant role in the diagnosis of benign and malignant diseases such as breast cancer and prostate cancer, in which the morphology measurements of histologic primitives serve as a basis of quantitative analyses. However, the complex nature of digital pathology data, such as diversity, present significant challenges for such segmentation task, which might lead to difficulties in feature extraction and instance segmentation. By converting complex pathology data into minable image features using artificial intelligence assisted pathologist''s analysis, it becomes possible to automatically extract quantitative information of individual primitives. Machine learning algorithms, in particular deep models, are emerging as leading tools in quantitative analyses of digital pathology. It has exhibited great power in feature learning with producing improved accuracy of various tasks. In this survey, we provide a comprehensive review of this fast-growing field. We briefly introduce popular deep models, including convolutional neural networks, fully convolutional networks, encoder-decoder architectures, recurrent neural networks, and generative adversarial networks, and summarize current deep learning achievements in various tasks, such as detection and segmentation. This study also presents the mathematical theory, key steps, main advantages and disadvantages, and performance analysis of deep learning algorithms, and interprets their formulations or modelings for specific tasks. In addition, we discuss the open challenges and potential trends of future research in pathology image segmentation using deep learning.