COVID-Net USPro: An Explainable Few-Shot Deep Prototypical Network for COVID-19 Screening Using Point-of-Care...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Disease Detection
• Point-of-Care Ultrasound Applications

Background:

• Coronavirus Disease 2019 (COVID-19) necessitates rapid screening methods to mitigate spread and healthcare system strain.
• Point-of-care ultrasound (POCUS) is a cost-effective imaging modality for assessing COVID-19 symptoms.
• Deep learning in medical imaging shows promise for accelerated COVID-19 diagnosis, but data scarcity is a challenge.

Purpose of the Study:

• To develop an explainable few-shot deep learning network for COVID-19 detection using limited ultrasound images.
• To address the challenge of limited annotated datasets in medical AI for pandemic scenarios.
• To create an AI tool that assists in rapid and accurate COVID-19 screening via POCUS.

Main Methods:

• Development of COVID-Net USPro, an explainable few-shot deep prototypical network.
• Training and evaluation using a small number of ultrasound images (few-shot learning).
• Incorporation of an explainability component to ensure decisions are based on clinically relevant patterns.

Main Results:

• COVID-Net USPro achieved high performance with only five training images: 99.55% accuracy, 99.93% recall, and 99.83% precision for COVID-19 positive cases.
• The network's decisions were validated by an experienced clinician, confirming reliance on relevant diagnostic patterns.
• The explainability feature confirmed that the AI identifies actual disease patterns.

Conclusions:

• Explainable AI, like COVID-Net USPro, can effectively detect COVID-19 from limited POCUS data.
• Network explainability and clinical validation are crucial for adopting AI in medical diagnostics.
• The open-sourced COVID-Net USPro promotes reproducibility and further research in AI for infectious diseases.