Clustering-based spatial analysis (CluSA) framework through graph neural network for chronic kidney disease...

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in medicine

Background:

• Machine learning in digital pathology aids in assessing kidney function and diagnosing chronic kidney disease (CKD).
• Current methods often require extensive expert annotations, which are time-consuming and impractical.
• There is a need for automated, efficient methods to analyze histopathology images for kidney disease assessment.

Purpose of the Study:

• To develop and validate a novel computational framework, clustering-based spatial analysis (CluSA), for kidney function assessment and CKD diagnosis.
• To leverage unsupervised learning to identify spatial relationships in kidney tissue, minimizing the need for expert annotations.
• To predict estimated glomerular filtration rate (eGFR) at biopsy and eGFR changes over one year using CluSA.

Main Methods:

• Developed CluSA, an unsupervised machine learning framework to learn spatial relationships between local visual patterns in kidney tissue.
• Utilized 107,471 histopathology images from 172 biopsy cores for clustering and deep learning model training.
• Incorporated spatial information by color-encoding clustered patterns and performing spatial analysis via graph neural networks.
• Employed a random forest classifier to predict CKD and eGFR using various feature groups.

Main Results:

• Achieved high accuracy in predicting eGFR at biopsy: 0.95 accuracy, 0.97 sensitivity, 0.90 specificity, and 0.96 AUC.
• Demonstrated strong performance in predicting one-year eGFR changes: 0.84 accuracy, 0.83 sensitivity, 0.85 specificity, and 0.85 AUC.
• The CluSA framework successfully identified novel predictors of kidney function and renal prognosis without expert annotation.

Conclusions:

• CluSA represents the first spatial analysis framework based on unsupervised machine learning for kidney pathology.
• The framework accurately classifies and predicts kidney function and prognosis without requiring expert annotations.
• CluSA offers a powerful tool for advancing kidney disease research and clinical diagnostics by identifying novel predictive features.