Fog-based deep learning framework for real-time pandemic screening in smart cities from multi-site tomographies

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Smart City Technologies

Background:

• Pandemic diseases pose significant global health challenges, straining international health infrastructure.
• Artificial Intelligence of Things (AIoT) offers potential for efficient pandemic control and diagnosis in smart cities.
• Integrating multi-source institutional data is a major hurdle for practical AIoT pandemic solutions.

Purpose of the Study:

• To present a novel framework for enhanced pandemic disease diagnosis through multi-site data fusion.
• To develop and evaluate a multi-decoder segmentation network for accurate COVID-19 lesion segmentation from cross-domain CT scans.
• To propose a fog computing (FC) technique for deploying AIoT models on edge nodes for real-time analysis.

Main Methods:

• Developed a novel multi-decoder segmentation network to process heterogeneous data from CT scans.
• Implemented a fog computing (FC) technique named PANDFOG for edge deployment of the segmentation network.
• Evaluated the network's performance on three public datasets for COVID-19 lesion segmentation.

Main Results:

• The multi-decoder segmentation network achieved an average Dice score of 89.9% and an average Surface Dice of 86.87%.
• The PANDFOG system successfully deployed the network on edge nodes, enabling practical, automated COVID-19 pneumonia analysis.
• Demonstrated accurate segmentation of infections from cross-domain CT scans, highlighting the model's robustness.

Conclusions:

• The proposed multi-decoder segmentation network effectively segments infections from diverse CT scan data.
• The PANDFOG system facilitates practical, low-latency deployment of AIoT diagnostic tools in clinical settings.
• This research enhances real-time patient monitoring and clinical decision-making for pandemic diseases like COVID-19.