Lung pneumonia severity scoring in chest X-ray images using transformers

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate lung pneumonia diagnosis and severity assessment via chest X-rays (CXR) require robust methods.
• Current severity quantification techniques often demand resource-intensive training.
• Efficient computational tools are needed for rapid COVID-19 identification and severity prediction.

Purpose of the Study:

• To develop an efficient and adaptable method for diagnosing lung pneumonia and assessing its severity using CXRs.
• To introduce a novel neural network model with a Vision Transformer (ViT) architecture for COVID-19 and lung disease severity quantification.
• To evaluate the model's generalizability across diverse CXR datasets.

Main Methods:

• Developed a novel image augmentation scheme and a Vision Transformer Regressor Infection Prediction (ViTReg-IP) model.
• Utilized a ViT architecture with a regression head for severity quantification.
• Evaluated model performance on multiple open-source chest radiograph datasets and compared it against existing deep learning methods.

Main Results:

• Achieved a minimum Mean Absolute Error (MAE) of 0.569 and 0.512 for geographic extent and lung opacity scores, respectively.
• Attained a maximum Pearson Correlation Coefficient (PC) of 0.923 and 0.855 for geographic extent and lung opacity scores.
• Demonstrated exceptional performance in severity quantification with robust generalizability and modest computational requirements.

Conclusions:

• The ViTReg-IP model offers an efficient and accurate solution for COVID-19 and lung disease severity assessment.
• The model exhibits strong generalizability across different datasets, making it adaptable for clinical use.
• Publicly available source code facilitates further research and application of the developed method.