Dr. Xiangmin Han, Postdoctoral Fellow, was awarded the fellowship from the China Postdoctoral Science Foundation

title: Research on Cross-dimensional High-order Correlation Computation Methods for 2D–3D Pathology Images

Current methods for pathology image analysis are predominantly based on 2D data, such as multiple instance learning and hypergraph computing. Due to the difficulty of acquiring 3D pathology images and their complex spatial structures and diverse tissue characteristics, existing 2D and 3D data analysis methods can hardly perform direct modeling and semantic computation on such data. Exploratory studies have attempted to introduce 3D convolutional neural networks and related techniques into 3D pathology image analysis, achieving performance superior to that on 2D pathology images. However, these methods still have notable limitations when processing 3D pathology images, particularly in recognizing and parsing complex three-dimensional spatial relationships among tissues and multi-scale features. They fail to sufficiently capture high-order spatial structures, which constrains the model’s generalization ability and robustness in complex pathological environments. Building on the applicant’s previous work on high-order correlation computation for 2D pathology images and preliminary explorations of 3D pathology images, this project aims to develop a high-order correlation computation method tailored to 3D pathology images. By constructing high-order relational structures in the 3D feature space, the proposed approach will enable high-order spatial analysis and representation learning of tumor tissues. In addition, this project will investigate adaptive high-order correlation collaborative computing methods for 2D–3D pathology images, so as to flexibly accommodate pathology image inputs of different dimensionalities, thereby providing strong theoretical and technical support for precise pathology image analysis and clinical decision-making.

面向2D-3D病理图的跨维度高阶关联计算方法研究

目前病理图分析的方法主要以2D数据分析为主，如多实例学习、超图计算。由于3D病理图像采集困难，且具有复杂的空间结构和多样的组织特征， 现有的2D及3D数据分析方法难以直接对其进行建模和语义计算。已有探索性的工作尝试将3D卷积神经网络等方法引入到3D病理图像分析中，取得 了较2D病理图像更优的性能。然而，此类方法在处理3D病理图像时仍然存在一定局限，尤其是在识别和解析组织间的复杂三维空间关系以及多尺度特征时， 无法充分捕捉高阶空间结构，限制了模型的泛化能力及在复杂病理环境中的鲁棒性。 基于申请人前期针对2D病理图的高阶关联计算基础以及对3D病理图的初步探索，本项目拟提出一种面向3D病理图像的高阶关联计算方法，通过在3D特征空 间中构建高阶关联结构，实现对肿瘤组织的高阶空间解析和表征学习。此外，本项目还将研究基于2D-3D病理图的自适应高阶关联协同计算方法，以满足不 同维度病理图输入的灵活应用需求，进而为病理图像的精准分析和临床决策提供强有力的理论及技术支撑。