Prediction of Tuberculosis From Lung Tissue Images of Diversity Outbred Mice Using Jump Knowledge Based Cell Graph...

Area of Science:

• Computational pathology
• Artificial intelligence in medical imaging
• Graph neural networks for disease detection

Background:

• Tuberculosis (TB) diagnosis relies on detecting acid-fast bacilli (AFB) in stained tissue samples.
• Whole Slide Imaging (WSI) enables digital analysis of pathology slides.
• Current deep learning methods for WSIs often use patch-wise analysis, potentially missing crucial spatial patterns like granulomas.

Purpose of the Study:

• To develop a novel deep learning approach for TB classification using cell graph modeling on WSIs.
• To capture complex cell interactions and tissue micro-architecture beyond patch-wise analysis.
• To improve diagnostic accuracy by integrating cell morphology and graph-based features.

Main Methods:

• A Convolutional Neural Network (CNN) was trained to segment AFBs and macrophage nuclei.
• Lung histology images were converted into cell graphs, with nodes representing cells and edges representing interactions.
• A cell graph-based jumping knowledge neural network (CG-JKNN) was developed, using biologically informed thresholds for edge determination.
• Integrated Gradients and SHAP were used for model interpretability.

Main Results:

• The CG-JKNN model achieved an F1 score of 0.9549 and an AUPRC of 0.9846.
• Extreme Gradient Boosting (XGBoost) with combined features performed best (F1: 0.9813, AUPRC: 0.9848).
• Identified significant features align with pathologist criteria, indicating clinical relevance.

Conclusions:

• The proposed cell graph-based approach effectively models cellular interactions for TB diagnosis from WSIs.
• Integrating graph and morphological features enhances classification accuracy.
• The method shows significant clinical applicability and potential for future advancements in TB diagnostics.