Learning Decision Ensemble using a Graph Neural Network for Comorbidity Aware Chest Radiograph Screening.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Graph Neural Networks

Background:

• Chest radiographs are crucial for screening cardiovascular, thoracic, and pulmonary conditions.
• Machine learning (ML) automates screening, aiding radiologists but often overlooks disease comorbidities.
• Current ML models, like deep convolutional neural networks (CNNs), have limitations due to ignoring disease interdependencies.

Purpose of the Study:

• To develop an advanced ML model that accounts for disease comorbidities in chest radiograph analysis.
• To improve the performance of automated screening systems by modeling disease dependencies.

Main Methods:

• A Graph Neural Network (GNN) was proposed to create ensemble predictions.
• The GNN models dependencies between different diseases for more accurate screening.
• Ensemble predictions were generated using the GNN with DenseNet121 architecture.

Main Results:

• The GNN-based ensemble method demonstrated improved performance over standard ensembling techniques.
• The model showed enhanced screening capabilities across various ensemble constructions.
• An average Area Under the Curve (AUC) of 0.821 was achieved for thirteen disease comorbidities using the GNN ensemble of DenseNet121.

Conclusions:

• The proposed GNN approach effectively models disease comorbidities, enhancing chest radiograph screening.
• This method offers a significant improvement over existing ML techniques for complex pulmonary diagnoses.
• The GNN ensemble shows promise for reducing radiologist workload and improving diagnostic accuracy.