Deep learning-based important weights-only transfer learning approach for COVID-19 CT-scan classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic necessitates rapid and accurate diagnostic tools.
• Traditional diagnostic methods can be time-consuming.
• Deep learning models offer potential for automated COVID-19 detection from CT scans, but face deployment challenges on devices with limited computational power.

Purpose of the Study:

• To propose an efficient transfer learning method for COVID-19 detection using CT scans.
• To adapt pre-trained deep learning models for deployment on resource-constrained, point-of-care devices.
• To reduce the computational and memory footprint of deep learning models for real-time inference.

Main Methods:

• A weights-only transfer learning approach was employed.
• Less important weight parameters in pre-trained models (VGG16, ResNet34) were pruned to create lighter models.
• The method focused on making models suitable for point-of-care devices with limited runtime resources.

Main Results:

• The pruned ResNet34 model achieved high accuracy (95.47%), sensitivity (0.9216), F-score (0.9567), and specificity (0.9942) on the SARS-CoV-2 CT-scan dataset.
• The optimized model demonstrated a significant reduction in computational load, with 41.96% fewer FLOPs and 20.64% fewer weight parameters.
• The proposed method effectively reduced runtime resource requirements for computationally intensive models.

Conclusions:

• The developed transfer learning method successfully reduces the resource demands of deep learning models for COVID-19 detection.
• The optimized models are suitable for deployment on point-of-care devices, facilitating real-time inference.
• This approach enhances the accessibility and practicality of AI-driven diagnostics in healthcare settings.