Interpretable classification of pathology whole-slide images using attention based context-aware graph convolutional...

Area of Science:

• Computational Pathology
• Medical Image Analysis
• Machine Learning

Background:

• Whole slide image (WSI) analysis for cancer detection is challenging due to large data size and need for context.
• Existing weakly supervised methods fail to capture inter-instance relationships, limiting predictive power and localization accuracy.

Purpose of the Study:

• To develop an interpretable classification model for WSIs that effectively utilizes context-aware features.
• To improve the accuracy and localization of lymph node metastasis detection in WSIs.

Main Methods:

• Proposing a bidirectional Attention-based Multiple Instance Learning Graph Convolutional Network (ABMIL-GCN).
• Hierarchically aggregating instance features using graph convolutional networks for global representation.
• Employing a topology-aware approach to model spatial correlations between image patches.

Main Results:

• Achieved average accuracy (ACC) of 90.89% and area under the curve (AUC) of 0.9149 on the Camelyon16 dataset.
• Demonstrated significant improvements over state-of-the-art methods, with ACC and AUC increases exceeding 7% and 4%, respectively.
• Validated the model's superiority in predicting slide labels and localizing regions of interest.

Conclusions:

• Context-aware graph convolutional networks outperform traditional weakly supervised methods by incorporating spatial relationships.
• The proposed ABMIL-GCN model addresses the accuracy-interpretability trade-off in computational pathology.
• This framework offers a new approach for computer-aided diagnosis and intelligent systems in clinical settings.