Related Experiment Video
Updated: Sep 22, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.0K
A Graph-Transformer for Whole Slide Image Classification
IEEE Transactions on Medical Imaging
|May 20, 2022
Summary
A new Graph-Transformer (GT) model, called GTP, improves whole slide image (WSI) analysis for cancer grading. This deep learning approach effectively distinguishes lung cancer subtypes from normal tissue, offering better accuracy than traditional patch-based methods.
Area of Science:
- Computational pathology
- Artificial intelligence in medicine
- Digital pathology image analysis
Background:
- Whole slide image (WSI) analysis is crucial for disease grading.
- Current patch-based deep learning methods for WSIs suffer from label noise and neglect WSI-level context.
- There is a need for advanced deep learning frameworks that capture global WSI information for accurate disease classification.
Purpose of the Study:
- To develop and evaluate a novel Graph-Transformer (GT) model, termed GTP, for WSI-level disease grading.
- To improve upon existing patch-based deep learning methods by integrating graph-based representations and vision transformers.
- To achieve accurate classification of lung cancer subtypes (adenocarcinoma and squamous cell carcinoma) and normal tissue from WSIs.
Main Methods:
- A dataset of 4,818 WSIs from CPTAC, NLST, and TCGA was utilized.
- A contrastive learning framework was developed using NLST data to create a feature extractor for WSI patch feature vectors.
- These feature vectors served as nodes for constructing the Graph-Transformer (GT) framework (GTP), fusing graph and vision transformer components.
Main Results:
- The GTP model achieved high performance in three-label classification (normal vs. LUAD vs. LSCC) with a mean accuracy of 91.2 ± 2.5% on CPTAC data via cross-validation.
- The model demonstrated strong generalization on external test data (TCGA), achieving a mean accuracy of 82.3 ± 1.0%.
- A novel graph-based saliency mapping technique, GraphCAM, was introduced for identifying class-discriminative regions within WSIs.
Conclusions:
- The Graph-Transformer (GT) model, GTP, represents an effective and interpretable deep learning framework for WSI-level classification.
- GTP overcomes limitations of patch-based methods by incorporating WSI-level context through graph representations.
- The developed GraphCAM technique enhances the interpretability of deep learning models in digital pathology.
