Core Concepts
Introducing an integrative graph-transformer framework for histopathology whole slide image representation and classification.
Abstract
Abstract:
Multiple instance learning (MIL) in weakly supervised histopathology WSI classification.
Graph-transformer framework for contextual information and global WSI representations.
Introduction:
Importance of digital pathology in cancer diagnosis.
Deep-learning based MIL for slide-level labeled WSIs.
Method:
Graph construction, backbone, and downstream process.
Graph-Transformer Integration Block for spatial relationships and global attention.
Experiments:
Testing on TCGA-NSCLC, TCGA-RCC, and BRIGHT datasets.
Superior performance over current state-of-the-art MIL methods.
Results:
Comparison with SOTA methods on accuracy and AUROC.
Ablation studies showcasing the effectiveness of the GTI block.
Conclusion:
Introduction of the IGT framework for histopathology WSI classification.
Integration of GCN and global attention for improved performance.
Stats
Existing attention-based MIL approaches often overlook contextual information and intrinsic spatial relationships between neighboring tissue tiles.
Extensive experiments on three publicly available WSI datasets show an improvement of 1.0% to 2.6% in accuracy and 0.7%-1.6% in AUROC.
Quotes
"Our IGT framework consistently outperforms existing state-of-the-art MIL methods."
"The self-attention mechanism in our GTI captures pairwise correlation across all instances and improves performance."