Exploiting Geometric Features via Hierarchical Graph Pyramid Transformer for Cancer Diagnosis Using Histopathological...

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Accurate cancer diagnosis relies heavily on pathology analysis of histopathological images.
• Existing deep learning methods often overlook crucial geometric features of cell distribution and tissue patterns.
• Geometric features are potent indicators for pathological image classification.

Purpose of the Study:

• To propose a novel Hierarchical Graph Pyramid Transformer (HGPT) for pathological image classification.
• To effectively exploit geometric representations of tissue distribution, a feature often ignored by current methods.
• To enhance the diagnostic capabilities for malignant tumors using deep learning.

Main Methods:

• Constructing a graph representation from pathological image morphological features.
• Utilizing a multi-head graph aggregator to learn geometric representations.
• Employing a transformer encoder to model long-range dependencies between image and graph data.
• Incorporating a locality feature enhancement block to improve 2D local representation.

Main Results:

• The HGPT method achieved superior classification outcomes on multiple cancer datasets (Kather-5K, MHIST, NCT-CRC-HE, GasHisSDB).
• Consistent performance improvements were observed for both binary and multi-category classification tasks.
• The approach demonstrated effectiveness in capturing and utilizing geometric features for improved accuracy.

Conclusions:

• The proposed HGPT offers an effective diagnostic tool for malignant tumors in clinical practice.
• Exploiting geometric features significantly enhances pathological image classification performance.
• This method represents a advancement in applying deep learning to computational pathology.