CCF-GNN: A Unified Model Aggregating Appearance, Microenvironment, and Topology for Pathology Image Classification

Area of Science:

• Computational pathology
• Graph Neural Networks
• Topological Data Analysis

Background:

• Pathology images offer rich data for cancer analysis, with cell topology being crucial for immunotherapy.
• Current deep learning methods often overlook topological features due to a lack of effective descriptors for cell distribution patterns.

Purpose of the Study:

• To develop a novel deep learning framework for pathology image classification that comprehensively learns cell appearance, microenvironment, and topology.
• To introduce a new topological descriptor, Cell Community Forest (CCF), for analyzing cell distribution patterns.

Main Methods:

• Designed Cell Community Forest (CCF) to represent hierarchical cell community formation.
• Proposed CCF-GNN, a Graph Neural Network model integrating cell-instance, cell-community, and image-level features.
• Employed a fine-to-coarse learning strategy for multi-level feature aggregation.

Main Results:

• CCF-GNN significantly outperforms existing methods in pathology image classification tasks.
• Demonstrated superior performance on H&E-stained and immunofluorescence images for disease grading across multiple cancer types.
• Validated the effectiveness of CCF as a geometric topological descriptor for tumor cells.

Conclusions:

• The proposed CCF-GNN framework effectively integrates multi-level heterogeneous features for pathology image analysis.
• This topological data analysis-based method advances the application of deep learning in computational pathology.
• The approach holds promise for improved cancer diagnosis and treatment strategy development.