Modeling long-range dependencies for weakly supervised disease classification and localization on chest X-ray.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Deep Learning for Diagnostics

Background:

• Computer-aided diagnosis (CAD) using chest X-rays (CXRs) is rapidly advancing due to deep learning, particularly convolutional neural networks (CNNs).
• CNNs' inherent locality limits their ability to capture long-range dependencies in medical images.
• Vision transformers (ViTs) address this but standard patch-based training in ViTs is insufficient for detailed medical image detection.

Purpose of the Study:

• To develop an improved CXR detection method that overcomes the limitations of traditional CNNs and standard ViTs.
• To effectively model both patch-wise and inter-patch dependencies for enhanced medical image analysis.
• To improve the accuracy and detail of thoracic disease classification and localization.

Main Methods:

• A novel CXR detection method integrating CNNs with an adaptive vision transformer (ViT) was proposed.
• A DenseNet architecture with a feature pyramid structure was employed to model long-range dependencies.
• The method was trained on the ChestX-ray14 dataset using weakly supervised learning due to global annotations.

Main Results:

• Achieved state-of-the-art performance in thoracic disease classification, with a mean area under the curve (AUC) of 0.829.
• Demonstrated significantly higher intersection of the union (IoU) scores for lesion localization compared to other methods.
• Visualizations confirmed more accurate and detailed predictions, with strong generalization ability shown on an external dataset.

Conclusions:

• The proposed integrated CNN-ViT method sets a new state-of-the-art for thoracic disease classification and weakly supervised localization.
• This advanced AI approach shows significant potential to aid clinicians in making informed diagnostic decisions.