LungNet-ViT: Efficient lung disease classification using a multistage vision transformer model from chest radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate classification of lung diseases from chest radiographs (CXR) is crucial for timely diagnosis and treatment.
• Existing methods may face challenges with dataset imbalance and feature extraction efficiency.

Purpose of the Study:

• To introduce and evaluate a Multistage-Vision Transformer (Multistage-ViT) model for precise lung disease classification using CXR images.
• To assess the model's performance on both imbalanced and balanced datasets.

Main Methods:

• Development of a hybrid Multistage-ViT model integrating backbone networks with the Vision Transformer (ViT) architecture.
• Utilizing deep feature extraction techniques for enhanced classification accuracy.
• Testing the model on a dataset comprising Normal, COVID-19, Viral Pneumonia, and Lung Opacity classes.

Main Results:

• The Multistage-ViT model achieved high classification accuracies: 99.93% on an imbalanced dataset and 99.97% on a balanced dataset.
• Hybrid configurations, particularly InceptionV3 combined with ViT, demonstrated superior performance over standalone models.
• The model effectively extracted deep features from CXR images, enhancing classifier performance.

Conclusions:

• The Multistage-ViT model offers superior accuracy and robustness for lung disease classification from CXR images.
• The integration of ViT architecture facilitates advanced deep feature extraction, improving diagnostic capabilities.
• This approach holds significant potential for advancing automated lung disease diagnosis.