基于深度全卷积网络的超高像素全切片病理图像分割和诊断技术研究

The analysis of histopathological images aided by artificial intelligent is significant for improving the accuracy and efficacy of pathological tissue diagnosis. In this research, to address the issues of the problems existing in the current analysis methods of gigapixel whole slide images (WSI), such as computationally expensive, partial label of samples and difficult of obtaining results of detailed lesions segmentation and accurate WSI classification, we mainly focus on the study of WSI analysis based on the framework of deep fully convention neural networks (FCNs), which can achieve multiple pathologic diagnoses from fine-grained pixel-level of lesions detection and region segmentation to macroscopic slide-level of histopathology classification. We first design a deep FCNs architecture based on the multi-scale features fusion and class activation mapping, and then a monitor network is applied to enhance the training of histopathology recognition model from the unbalanced distribution and noisy labels of samples, so that we can generate lesion probability hot maps efficiently and effectively. And then the fine-grained segmentation results of smooth lesion region border are generated by the Encoder-Decoder FCNs based on the weighted feedback with partial labels, which are further refined by combing the post-processing methods including conditional random fields and super-pixels. Finally, the features of WSI are extracted adaptively from most representative pathological patches and an end to end pathological tissue diagnosis is generated based on the spatial pyramid pooling deep neural network. In summary, the implementation of this project could provide new clues for the research and application of artificial intelligence aided histopathology images diagnosis.

人工智能辅助组织病理图像分析对于提高病理诊断准确性和效率具有重要意义。本项目针对当前超高像素全切片病理图像分析中存在的计算量大、样本标记不完全、难以获取病变区域精细分割和精确病理诊断等问题，开展基于深度全卷积网络框架的全切片病理图像智能分析研究，实现从微观细节的像素级病变区域感知、边界分割到宏观层面的全切片组织病理定性诊断。首先，研究基于多尺度特征融合和类激活映射的深度全卷积网络，并结合监督网络训练加强在有损标记和不平衡样本下识别模型的泛化能力，高效精确生成病变区域热点图；其次，研究基于部分标记加权反馈的编解码全卷积分割网络，并融合条件随机场和超像素等后处理方法进行修正，得到病变区域完整平滑的分割边界；最后，研究基于自适应显著病变区域特征提取方法，并构建空间金字塔池化的深度网络模型，实现端到端的全切片病理组织诊断。通过本项目的实施为探索和应用人工智能辅助病理图像诊断提供新的思路。

人工智能辅助组织病理图像分析对于提高病理诊断准确性和效率具有重要意义。在病理分析中，组织病理图像特征冗余性高，组成组织复杂，训练样本不足，因此难以构建精确病理诊断模型。同时，全切片图像往往高达上亿像素，存在大量干扰和噪声等情况，单纯的利用深度全卷积网络难以挖掘高维全切片图像中细微的病灶差异。最后，超高像素病理图像面临样本标记不完全、深度卷积网络计算量大、在有损标记以及样本多样性等问题，难以获取病变区域以进行精细识别和病理定性诊断。..针对上述病理诊断难点问题，本项目开展了基于深度卷积网络框架的病理图像自动化分析和诊断方法研究。首先，提出了基于自监督对比学习的深度卷积网络图像分析框架，实现了基于高维图像成对学习、多标签对比学习、临床分数对比学习、生成对抗学习等模型和方法，提高了在小样本不平衡分布下的超高像素病理图像分析识别精度，为构建弱语义分割模型提供基础；其次，提出了面向高维医学图像的多注意力特征提取机制，实现了多尺度空间注意力校正、多尺度注意力融合等方法，提高了在超高维度下病理图像的关键目标特征提取能力，为超高像素全切片细粒度精确诊断提供基础；最后，提出了基于不完备信息与多模态信息融合的诊断决策技术，提高了在不完备标注以及多模态信息辅助下的病理诊断能力。..经过三年的研究，本项目共计发表SCI/EI国际学术论文12篇，其中包括CCF推荐期刊和会议论文6篇（NCAA,WWW，IJCNN,ICPR,IEEE-SMC），中科院分区2区及以上期刊论文3篇；已申请国内/国际发明专利3项，其中获授权发明专利1项。已发表的论文均标注了该项目的支持。