Deep learning representations to support COVID-19 diagnosis on CT slices

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Coronavirus disease 2019 (COVID-19) presents a global health challenge.
• Computed tomography (CT) scan analysis aids in COVID-19 diagnosis but can be subjective.
• Automated analysis tools are needed to improve objectivity and efficiency.

Purpose of the Study:

• To develop and evaluate deep learning models for automated COVID-19 detection using thoracic CT scans.
• To assess the efficacy of transfer learning in extracting relevant features for COVID-19 classification.
• To compare the performance of end-to-end deep learning with traditional classifiers using deep features.

Main Methods:

• Utilized two datasets: SARS-CoV-2 CT Scan (Set-1) and FOSCAL clinic's dataset (Set-2).
• Applied transfer learning by adapting pre-trained deep learning models from natural image domains.
• Performed classification using an end-to-end deep learning approach and by feeding deep features into Random Forest and Support Vector Machine classifiers.

Main Results:

• The end-to-end deep learning model achieved high accuracy: 92.33% for Set-1 and 96.99% for Set-2.
• Deep feature embedding with Support Vector Machine also demonstrated strong performance: 91.40% accuracy for Set-1 and 96.00% for Set-2.
• Both methods showed high precision, indicating reliable identification of COVID-19 cases.

Conclusions:

• Deep learning representations show excellent performance in identifying COVID-19 from CT scans.
• These models effectively characterize COVID-19 radiological patterns, supporting diagnostic accuracy.
• The developed approach holds potential for clinical application in COVID-19 diagnosis.