Contrastive Cross-Site Learning With Redesigned Net for COVID-19 CT Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic necessitates rapid and accurate diagnostic tools.
• Automated identification of COVID-19 from CT images can aid clinical diagnosis and reduce radiologist workload.
• Aggregating diverse datasets is essential for developing robust machine learning models but faces challenges due to distribution discrepancies.

Purpose of the Study:

• To propose a novel joint learning framework for accurate COVID-19 identification from heterogeneous CT image datasets.
• To address the distribution discrepancy and cross-site domain shift inherent in multi-site medical data.
• To enhance the generalizability and performance of COVID-19 detection models.

Main Methods:

• Redesigned COVID-Net architecture and learning strategy for improved accuracy and efficiency.
• Implemented separate feature normalization in latent space to mitigate cross-site domain shift.
• Utilized a contrastive training objective to promote domain-invariant semantic embeddings.

Main Results:

• The proposed framework demonstrated consistent performance improvements on two public COVID-19 CT datasets.
• Achieved significant improvements in Area Under the Curve (AUC) compared to the original COVID-Net (12.16% and 14.23%).
• Outperformed existing state-of-the-art multi-site learning methods for COVID-19 identification.

Conclusions:

• The joint learning framework effectively addresses challenges in learning from heterogeneous medical imaging datasets.
• The method offers a robust solution for accurate and generalizable COVID-19 identification using CT scans.
• This approach has the potential to significantly support clinical decision-making in pandemic scenarios.