Big data analysis for Covid-19 in hospital information systems

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic necessitates efficient diagnostic tools for clinical decision-making.
• Automated analysis of CT images can aid in rapid COVID-19 identification and reduce radiologist workload.
• Heterogeneous medical datasets present challenges for developing robust machine learning models due to distribution discrepancies.

Purpose of the Study:

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

Main Methods:

• A novel deep learning joint framework incorporating redesigned COVID-Net architecture and learning strategies.
• Independent feature normalization in latent space to improve prediction accuracy and learning efficiency.
• Contrastive training objective to enhance domain invariance of semantic embeddings and boost classification performance.

Main Results:

• The proposed method significantly improves performance on two large-scale public COVID-19 CT datasets.
• Outperformed original COVID-Net by 13.27% and 15.15% in AUC on respective datasets.
• Exceeded the performance of existing state-of-the-art multi-site learning methods.

Conclusions:

• The developed deep learning framework effectively handles heterogeneous datasets for accurate COVID-19 identification.
• The approach demonstrates improved accuracy, learning efficiency, and domain invariance.
• This method offers a robust solution for multi-site COVID-19 diagnosis using CT imaging.