Real-Time Fluorescence-Based COVID-19 Diagnosis Using a Lightweight Deep Learning System.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Deep learning models for COVID-19 diagnosis using CT/X-ray images face cost, time, and radiation limitations.
• Real-time COVID-19 diagnosis necessitates lightweight deep learning models suitable for embedded systems.

Purpose of the Study:

• To propose and validate a lightweight deep learning model for COVID-19 diagnosis using fluorescence images.
• To demonstrate the model's feasibility for real-time diagnosis on low-power edge devices.

Main Methods:

• Fluorescence images were preprocessed (Gray Scale, CLAHE, Z-Score normalization) to address data imbalance.
• ResNet152 and VGG13 architectures were selected based on initial accuracy and then pruned using layer-wise importance calculation.
• Model lightweighting was achieved by pruning less important layers to reduce size and parameters.

Main Results:

• Pruned VGG13 maintained accuracy, reducing size by 18.9 MB and parameters by 4.2 M.
• Pruned ResNet152 improved accuracy by 1%, reducing size by 161.5 MB and parameters by 40.22 M.
• The optimized model achieved 7.69 FPS on an NPU, demonstrating real-time diagnostic capability.

Conclusions:

• Lightweight deep learning models can enable real-time COVID-19 diagnosis on resource-constrained edge devices.
• Fluorescence imaging combined with optimized deep learning offers a viable alternative to traditional methods.