Automated COVID-19 Grading With Convolutional Neural Networks in Computed Tomography Scans: A Systematic Comparison

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Radiologist workload for COVID-19 CT assessment is high during pandemics.
• Automated COVID-19 classification using CNNs is an active research area.
• Previous studies lacked systematic comparison of CNN components.

Purpose of the Study:

• To systematically evaluate 3-D CNNs versus 2-D CNNs for COVID-19 classification from CT images.
• To identify optimal components for improving COVID-19 detection algorithms.
• To establish a benchmark for fair comparison in future research.

Main Methods:

• Systematic comparison of seven common CNN architectures (DenseNet, Inception, ResNet variants) using 2-D and 3-D approaches.
• Investigated effects of pre-trained weights, lesion maps, and continuous output prediction on the best-performing architecture.
• Utilized a test set of 105 CT scans and a public dataset of 742 CT scans.

Main Results:

• A 3-D DenseNet-201 architecture achieved an Area Under the Curve (AUC) of 0.930 on the internal test set.
• The same model reached an AUC of 0.919 on the public dataset, outperforming a 2-D CNN.
• Optimized components led to substantial improvements in COVID-19 classification performance.

Conclusions:

• 3-D CNNs offer significant performance benefits over 2-D CNNs for COVID-19 classification from CT scans.
• Systematic evaluation and component optimization are crucial for developing effective AI tools in medical imaging.
• The study provides valuable insights for developing robust AI-powered COVID-19 diagnostic systems.