EffViT-COVID: A dual-path network for COVID-19 percentage estimation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic caused a global health crisis, necessitating advanced diagnostic tools.
• Accurate estimation of COVID-19 infection percentage is crucial for patient management and treatment planning.
• Existing vision-based methods for COVID-19 prediction lack focus on infection percentage estimation.

Purpose of the Study:

• To propose a novel deep learning framework for estimating COVID-19 infection percentage.
• To enhance the precision of COVID-19 diagnosis by quantifying infection severity.
• To develop a robust tool for assessing lung involvement in COVID-19 patients.

Main Methods:

• A deep learning network integrating features from vision transformers and EfficientNet-B7 (CNN).
• Feature fusion technique to create an information-rich vector for precise estimation.
• Evaluation on the Per-COVID-19 dataset using slice-level metrics (PC, MAE, RMSE) and 5-fold cross-validation.

Main Results:

• The proposed network achieved state-of-the-art performance in COVID-19 infection percentage estimation.
• Achieved high Pearson correlation coefficient (PC), low Mean Absolute Error (MAE), and Root Mean Square Error (RMSE).
• Demonstrated a low overall average difference between actual and predicted infection percentages, indicating high accuracy.

Conclusions:

• The developed deep learning framework is robust and efficient for estimating COVID-19 infection percentage.
• The fusion of vision transformers and CNNs significantly improves estimation accuracy.
• This work contributes a valuable tool for objective assessment of COVID-19 severity using CT imaging.