Multi-scale nested graph transformer with graph operations: Advancing high-resolution chest x-ray classification.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate classification of high-resolution chest x-ray (CXR) images is crucial for diagnosing lung conditions and identifying small lesions.
• Traditional deep learning models struggle with balancing local detail and global context, especially with limited data and high computational costs.

Purpose of the Study:

• Introduce a Multi-scale Nested Graph Transformer (MNGT) to enhance CXR classification accuracy and computational efficiency.
• Improve model generalization in data-constrained scenarios for high-resolution CXR analysis.

Main Methods:

• Employ a multi-scale nested architecture segmenting CXR images into hierarchical patches for local-to-global feature extraction.
• Utilize cross-attention fusion between high-resolution and low-resolution images to improve lesion discriminability.
• Incorporate graph pooling for computational efficiency and inductive bias integration (graph convolution, adaptive receptive fields) to mitigate overfitting and improve generalization.

Main Results:

• The MNGT architecture demonstrated superior performance over other models in accuracy and F1-score on three high-resolution CXR datasets.
• Ablation studies confirmed the architectural efficiency of the MNGT model.
• Experimental results highlight the model's effectiveness in handling high-resolution medical imaging challenges.

Conclusions:

• MNGT offers an efficient and robust solution for high-resolution CXR classification, excelling in accuracy and generalization.
• The framework effectively addresses computational bottlenecks in medical imaging, paving the way for clinical computer-aided diagnosis deployment.