Efficient and visualizable convolutional neural networks for COVID-19 classification using Chest CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deep learning shows promise for diagnosing coronavirus disease 2019 (COVID-19).
• Comparing deep learning models for COVID-19 diagnosis is challenging due to varied data and acquisition methods.
• Accurate characterization of COVID-19 patients is crucial amidst rising case numbers.

Purpose of the Study:

• To design, evaluate, and compare 20 convolutional neural networks for classifying COVID-19 positive, healthy, and other pulmonary infections using chest CT scans.
• To investigate the performance of the EfficientNet family for COVID-19 diagnosis.
• To utilize intermediate activation maps for visualizing model performance and interpretability.

Main Methods:

• Trained and evaluated 20 convolutional neural networks on 4173 chest CT images from "A COVID multiclass dataset of CT scans."
• Classified patients into three categories: COVID-19 positive, healthy, and other pulmonary infections.
• Employed EfficientNet models and visualized performance using intermediate activation maps and Gradient-weighted Class Activation Mappings.

Main Results:

• EfficientNet-B5 achieved the highest performance with an F1 score of 0.9769 ± 0.0046 and accuracy of 0.9759 ± 0.0048 on the multiclass dataset.
• On a 2-class dataset, EfficientNet-B5 demonstrated high accuracy (0.9845 ± 0.0109) and F1 score (0.9599 ± 0.0251).
• Activation maps provided interpretable evidence of the model's focus on opacities and consolidations.

Conclusions:

• EfficientNet-B5 is a highly accurate and efficient model for diagnosing COVID-19 and other pulmonary infections from chest CT scans.
• The AI-assisted radiology tool offers rapid (under 0.5s) and scalable diagnostic capabilities.
• Intermediate activation maps enhance the interpretability of deep learning models in medical imaging.