融合深度全卷积网络和稀疏字典的病理细胞核分割研究

The morphology and distribution of nuclei are the main basis for precision diagnosis and prognosis evaluation of breast cancer. Automated nucleus segmentation remains to be a challenging problem due to the complex nature of hematoxylin-eosin (HE) stained histopathology images. Single segmentation algorithms often have negative effects on large-scale images. This project present and implement an integrated and learning-based framework for automatic nucleus segmentation in HE stained breast histopathology images. The framework is mainly composed of three parts: modeling of the multi-level context information and texture information between inside and outside the nucleus, and construct a deep fully convolutional network detection model for obtaining initial positions of nuclei; analysis of the shape priors, and construct a sparse dictionary learning priors model for preserving boundary details; analysis of the local spatial constraints, and construct a boundary evolution model based on deformable theory for avoiding over segmentation. At last a nucleus segmentation application system in HE stained histopathology images will be developed, and the above theoretical research results will be applied in the system. The research of this project will have important theoretical significance and application value for promoting the establishment of computer-aided diagnosis and prognosis prediction system.

细胞核的形态和结构分布是乳腺癌精准诊断和预后评估主要依据。由于苏木素-伊红染色（Hematoxylin-Eosin,HE)组织病理图像的复杂性，细胞核自动分割是一个挑战性的难题。单一分割算法在大范围图像内往往产生负面效果，本项目提出并实现一种集成的基于学习的HE乳腺病理图像细胞核自动分割框架，该框架主要由三个部分组成：建模图像中多层次上下文环境信息和细胞核内外纹理差异信息，构建一种基于深度全卷积神经网络的检测模型初始细胞核位置；分析病理形状先验构建一种基于稀疏字典学习的形状先验模型保持边界形状细节；分析局部空间约束构建基于形变理论的细胞核边界演化模型避免过度分割。最后研制一个针对HE染色病理图像的细胞核分割系统，并将上述理论研究成果应用于该系统。该课题的研究对促进计算机辅助病理诊断和预后预测系统的建立有重要的理论意义和应用价值。

细胞核的形态和结构分布是乳腺癌精准诊断和预后评估的主要依据。由于苏木青-伊红染色组织病理图像的复杂性，细胞核自动分割是一个挑战性的难题。传统分割算法在大范围内往往产生负面效果。近年来，随着深度学习在医学图像处理领域的巨大成就，提出了大量基于深度学习的细胞核分割算法，然而多数算法关注于分割精度的提高，忽略了临床实践中资源受限，执行速度要求高等问题。本研究中我们设计基于深度卷积网络的细胞核实例分割算法，针对大量的粘连和聚簇细胞，提出基于SSD模型的细胞核边框检测分支并将其应用于特征图的裁剪，以引导细胞核的实例分割，提高细胞核分割的精度。针对临床中计算资源受限，计算速度要求高等问题，我们进一步优化网络结构，提出基于UNET的统一实例分割框架，进一步提升模型的执行速度；最后，在病理图像分割算法分析工作的基础上，对课题研究内容进行延伸拓展到新冠CT图像，结肠镜图像和黑色素皮肤图像等医学图像分割，减少模型参数并提高分割精度。该研究对推动数字病理算法在临床上的应用，促进计算机辅助病理诊断和预后预测的建立有重要的理论意义和应用价值。