A vision transformer for emphysema classification using CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Emphysema involves lung alveoli destruction, classified into centrilobular (CLE), panlobular (PLE), and paraseptal (PSE) subtypes based on CT appearance.
• Automated classification of emphysema subtypes offers precise quantification of lung destruction patterns.

Purpose of the Study:

• To develop and evaluate a vision transformer (ViT) model for automated classification of emphysema subtypes using CT images.

Main Methods:

• A ViT model was proposed, processing CT image patches through embedding, transformer encoder blocks, and a softmax layer for classification.
• Transformer encoder blocks were pre-trained on ImageNet and fine-tuned to address limited medical imaging data.

Main Results:

• The pre-trained ViT model achieved an average accuracy of 95.95% on an internal dataset, outperforming other deep learning models like AlexNet and ResNet.
• Performance on public datasets was comparable or superior to existing methods, demonstrating the efficacy of pre-training.

Conclusions:

• The proposed ViT model accurately classifies emphysema subtypes from CT images.
• This method shows potential for effective computer-aided diagnosis of emphysema and can be extended to other medical imaging applications.