Multi-task vision transformer using low-level chest X-ray feature corpus for COVID-19 diagnosis and severity...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Developing accurate algorithms for COVID-19 diagnosis and severity quantification from Chest X-rays (CXR) is challenging due to limited curated datasets during the pandemic.
• Abundant CXR data with various findings exists, presenting an opportunity for machine learning approaches.
• Standard Vision Transformer (ViT) architectures may not be optimal for CXR analysis due to their feature embedding methods.

Purpose of the Study:

• To propose a novel Multi-task Vision Transformer (ViT) model for diagnosing and quantifying COVID-19 severity using CXR.
• To leverage abundant, unlabeled CXR data by incorporating a backbone network trained on common radiological findings.
• To enhance the feature embedding process for CXR analysis within the ViT architecture.

Main Methods:

• A backbone network was trained on large public datasets to extract common CXR findings (e.g., consolidation, opacity, edema).
• Features extracted by the backbone network were used as input corpora for a versatile Transformer model.
• The Multi-task ViT model was designed for both COVID-19 diagnosis and severity quantification.
• Model performance and generalization were evaluated on diverse external test datasets from different institutions.

Main Results:

• The proposed Multi-task ViT achieved state-of-the-art performance in both COVID-19 diagnosis and severity quantification.
• The model demonstrated outstanding generalization capabilities across various external datasets.
• The approach effectively utilized low-level CXR features for improved diagnostic accuracy.

Conclusions:

• The novel Multi-task ViT model offers a robust solution for COVID-19 diagnosis and severity assessment using CXR.
• The model's strong generalization capability is crucial for its potential widespread clinical deployment.
• Leveraging common CXR findings through a pre-trained backbone enhances the performance of Transformer-based models for specific medical imaging tasks.