Advancing COVID-19 differentiation with a robust preprocessing and integration of multi-institutional open-repository...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology and diagnostic imaging

Background:

• The COVID-19 pandemic accelerated AI research in medical diagnostics.
• Medical imaging, including CT scans, is crucial for COVID-19 diagnosis.
• Open imaging data repositories facilitate global research but present quality challenges.

Purpose of the Study:

• To develop and evaluate AI models for automated COVID-19 diagnosis using medical imaging.
• To assess the impact of data preprocessing on model performance with heterogeneous datasets.
• To compare AI model performance against existing diagnostic methods.

Main Methods:

• Implementation of a custom U-Net model for image segmentation.
• Development of a VGG-19 based transfer learning model for disease differentiation.
• Training and validation on a heterogeneous dataset of tomographic slices.
• Comparison of model performance with and without segmentation preprocessing.

Main Results:

• The custom U-Net model achieved a Dice similarity coefficient of 99.6%.
• The VGG-19 model demonstrated an Area Under the Curve of 96.1% for COVID-19 vs. pneumonia differentiation.
• Performance significantly improved compared to a baseline model without segmentation.

Conclusions:

• Robust preprocessing is vital for AI model performance in heterogeneous medical imaging datasets.
• The developed AI models show high diagnostic potential for COVID-19.
• AI-powered imaging analysis offers a valuable tool for infectious disease diagnosis.