Classifier Fusion for Detection of COVID-19 from CT Scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnosis

Background:

• Coronavirus disease (COVID-19), caused by SARS-CoV-2, is highly infectious, necessitating rapid and accurate diagnostic methods.
• Limited sensitivity and availability of testing kits highlight the need for alternative screening tools like Computer Tomography (CT) scans.
• Deep learning models show promise for COVID-19 detection but require large labeled datasets, which are scarce for this disease.

Purpose of the Study:

• To investigate the efficacy of pre-trained network architectures using transfer learning for COVID-19 detection in medical imaging datasets, particularly under data-limited conditions.
• To evaluate different variants of the ResNet model (ResNet18, ResNet50, ResNet101) for their performance in identifying COVID-19 from CT scans.
• To enhance detection accuracy by exploring model activations with traditional classifiers and proposing a classifier fusion strategy.

Main Methods:

• Utilized transfer learning with pre-trained ResNet architectures (ResNet18, ResNet50, ResNet101) for COVID-19 detection from CT scans.
• Evaluated the performance of individual ResNet models, focusing on recall and F1-score.
• Extracted features from different layers of the best-performing model and employed Support Vector Machine (SVM), Logistic Regression, and K-Nearest Neighbors (KNN) classifiers.
• Implemented a majority voting-based classifier fusion strategy to combine predictions from multiple classifiers.

Main Results:

• The transfer-learned ResNet50 model demonstrated superior performance, achieving a recall of 98.80% and an F1-score of 98.41%.
• Integrating learned image features with SVM, logistic regression, and KNN classifiers showed potential for improved detection.
• The proposed classifier fusion strategy further enhanced diagnostic accuracy, leading to a recall of 99.20% and an F1-score of 99.40%.

Conclusions:

• Transfer learning with pre-trained ResNet models, particularly ResNet50, is effective for COVID-19 detection using CT scans, even with limited datasets.
• Combining deep learning features with traditional classifiers and employing a fusion strategy significantly improves detection performance.
• The developed approach offers a promising, accurate, and efficient method for screening COVID-19 in clinical settings.