基于深度学习的多中心HE染色乳腺病理图像细胞核分割与识别方法研究

HE染色组织病理图像细胞核分割与识别是医学图像分析领域的研究热点和前沿。但细胞核尺寸大小不一，常呈现轮廓断裂、模糊的现象，容易导致细胞核漏检测和误分割；病理图像标注成本高、不同中心的图像存在较大差异性，单中心数据训练的模型泛化能力不佳。本项目围绕上述问题展开深入研究，提出新型的图像特征提取和学习方法：（1）将细胞核的初始形状集转化为卷积核，作为可学习组件嵌入深度网络中，构建形状先验项和形状相似度量项，进行细胞核形状特征的提取和优化，指导细胞核分割；（2）探索一种面向多中心细胞核分割的深度元学习方法：提出多尺度深度全卷积语义分割网络作为基学习器提取细胞核特征，元学习器指导基学习器寻找较优的初始化和参数更新方向；（3）研究基于人工特征、深度特征以及二者结合的集成分类方法，充分利用细胞核图像多维度特征。本项目的研究内容对于病理图像定量分析技术水平的提高、精准医疗诊断水平的提高都具有重要意义。

Nuclei segmentation and recognition of HE stained histopathological images is the hot research topic and frontier of medical image analysis. However, the size of the nuclei is different and the outline of nuclei is often broken and blurred, which easily leads to the missed detection and incorrect segmentation of the nuclei. In addition, the cost of pathological image annotation is high; the images in multi-center are quite different, and the generalization of the model that trained in single center data is poor. This project focuses on the above problems and proposes new image feature extraction and learning methods. First, the initial shape set of nuclei are transformed into convolution kernels, which can be embedded in the deep network as a learnable component, constructing shape prior term and shape similarity measurement term, extracting and optimizing the shape features of nuclei, and guiding the nuclei segmentation. Second, a meta-learning segmentation method is explored for multi-center nuclei segmentation. A multi-scale deep convolution semantic network is proposed as the base-learner to extract features of nuclei, and the meta-learner guides the base-learner to find the better model initialization and parameter update direction. Third, ensemble classification method is proposed to integrate classifier based on the hand-crafted feature, deep learning feature and the combination of the two, and make full use of the multi-dimensional features of the image. The research content of this project is of great significance for the improvement of quantitative analysis technology of pathological images and the improvement of precision medical diagnosis.