COVID-19 Identification from Low-Quality Computed Tomography Using a Modified Enhanced Super-Resolution Generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Computed Tomography (CT) is crucial for COVID-19 detection but often suffers from low image quality.
• High mortality rates and expert overload necessitate automated diagnostic tools for COVID-19 screening.
• Existing methods struggle with image resolution, impacting diagnostic accuracy.

Purpose of the Study:

• To enhance the performance of COVID-19 identification using CT images.
• To address the challenge of low quality and resolution in CT scans for COVID-19 detection.
• To introduce a novel deep learning framework for accurate and robust COVID-19 screening.

Main Methods:

• Developed a modified enhanced super-resolution generative adversarial network (SRGAN) to improve CT image resolution.
• Incorporated a Siamese capsule network to extract distinct features for COVID-19 identification, avoiding increased network complexity.
• Validated the model on the publicly available COVID-CT dataset.

Main Results:

• The proposed model achieved high performance metrics: 97.92% accuracy, 98.85% sensitivity, 97.21% specificity, 98.03% AUC, 98.44% precision, and 97.52% F1 score.
• Demonstrated state-of-the-art performance in COVID-19 identification compared to existing methods.
• The technique effectively handles low-quality CT images, proving robust for screening.

Conclusions:

• The proposed deep learning model offers an effective, accurate, and robust solution for COVID-19 screening using CT images.
• This novel framework, combining enhanced resolution and distinct feature extraction, is influential for diagnosing COVID-19 and related conditions.
• The approach is particularly valuable given the limited availability of medical imaging datasets for training.