Global attention based GNN with Bayesian collaborative learning for glomerular lesion recognition.

Area of Science:

• Computational pathology
• Renal pathology
• Machine learning for medical imaging

Background:

• Glomerular lesions are key indicators of kidney disease progression.
• Pathological diagnosis of these lesions is definitive but labor-intensive.
• Current deep learning methods struggle with complex spatial relationships in pathology images.

Purpose of the Study:

• To develop an advanced deep learning model for accurate glomerular lesion recognition.
• To overcome limitations of Euclidean space-based methods in pathology image analysis.
• To enhance feature extraction and classification for renal disease diagnosis.

Main Methods:

• Proposed a graph neural network (GNN) with global attention pooling (GAP) for semantic feature extraction.
• Incorporated Bayesian collaborative learning (BCL) for improved node feature fusion.
• Implemented a soft classification head to address semantic ambiguity in classification.

Main Results:

• Achieved high F1 scores (81.37%-98.68%) on four private glomerular datasets.
• Outperformed existing models in glomerular lesion recognition tasks.
• Demonstrated superior performance on a public breast cancer dataset (85.61% F1 score).

Conclusions:

• The GNN model enables precise recognition of glomerular lesions and aids kidney disease diagnosis.
• The framework is adaptable for various pathology image classification challenges.
• This approach offers a robust computational tool for renal pathology and disease diagnostics.