Label correlation transformer for automated chest X-ray diagnosis with reliable interpretability

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computer-aided diagnosis (CAD) of chest X-ray (CXR) images aids radiologists by reducing workload and inter-observer variability in disease screening.
• Current deep learning methods for multi-label CXR classification face challenges with accuracy and interpretability.

Purpose of the Study:

• To propose a novel transformer-based deep learning model for automated CXR diagnosis.
• To enhance classification performance and provide reliable interpretability for diagnostic tasks.

Main Methods:

• A novel transformer architecture was developed to capture global and local image information and label correlations.
• A new loss function was introduced to identify correlations between labels in CXR images.
• Attention heatmaps were generated for model interpretability, comparing model focus with physician-labeled pathogenic regions.

Main Results:

• The proposed model achieved a mean AUC of 0.831 on the Chest X-ray 14 dataset and 0.875 on the PadChest dataset.
• Performance surpassed existing state-of-the-art methods in CXR multi-label classification.
• Attention heatmaps demonstrated the model's ability to focus on relevant pathogenic regions.

Conclusions:

• The transformer-based model significantly improves the performance and interpretability of automated CXR diagnosis.
• The findings offer new evidence and methods for clinical applications in automated disease screening.
• The model provides reliable interpretability by correlating attention heatmaps with true pathogenic regions.