Hierarchical graph representations in digital pathology.

Area of Science:

• Computational pathology
• Graph-based machine learning
• Biomedical image analysis

Background:

• Accurate cancer diagnosis, prognosis, and treatment response prediction rely on understanding tissue structure and cell distribution.
• Current cell-graph methods capture cell-microenvironment interactions but often miss broader tissue architecture.
• A comprehensive tissue representation is crucial for advancing computer-aided cancer patient care.

Purpose of the Study:

• To develop a novel multi-level hierarchical entity-graph representation for tissue specimens.
• To create a hierarchical graph neural network (HACT-Net) for analyzing these representations.
• To improve the accuracy of mapping tissue structure to clinical functionality in cancer.

Main Methods:

• Proposed a multi-level hierarchical entity-graph to encode histological entities and their interactions at multiple scales.
• Developed HierArchical Cell-to-Tissue (HACT) graph representations from histology images, including cells and tissue regions.
• Designed HACT-Net, a message-passing graph neural network, to classify HACT representations.

Main Results:

• Introduced the BReAst Carcinoma Subtyping (BRACS) dataset for benchmarking.
• HACT-Net demonstrated superior classification performance compared to existing methods and individual pathologists.
• Ablation studies confirmed the effectiveness of the proposed hierarchical approach.

Conclusions:

• The hierarchical entity-graph representation effectively captures complex tissue structures.
• HACT-Net offers a powerful tool for computer-aided diagnosis in oncology.
• This approach advances the field of computational pathology for improved cancer patient outcomes.