Radiologists versus Deep Convolutional Neural Networks: A Comparative Study for Diagnosing COVID-19

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Reverse transcriptase polymerase chain reaction (RT-PCR) is the standard for COVID-19 diagnosis but suffers from low sensitivity, necessitating repeat testing.
• Chest computed tomography (CT) is an effective tool for diagnosing COVID-19 due to its direct visualization of lung involvement.

Purpose of the Study:

• To investigate the efficacy of pre-trained deep learning models in diagnosing COVID-19 from chest CT images.
• To compare the diagnostic performance of deep learning models against thoracic radiologists.

Main Methods:

• A dataset of 3000 chest CT images was utilized to train three pre-trained deep learning models: ResNet-18, ResNet-50, and DenseNet-201.
• The models and two thoracic radiologists were evaluated on a separate test set of 250 images.

Main Results:

• Deep learning models achieved higher accuracy (97.8%), sensitivity (98.1%), specificity (97.3%), precision (98.4%), and F1-score (98.25%) in classifying COVID-19 positive cases.
• The employed deep neural networks outperformed two thoracic radiologists in the diagnostic evaluation.

Conclusions:

• Pre-trained deep learning models demonstrate significant potential for accurate and efficient COVID-19 diagnosis using chest CT scans.
• AI-powered diagnostic tools can augment or potentially surpass human expert performance in identifying COVID-19 from medical imaging.