A self-supervised COVID-19 CT recognition system with multiple regularizations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Machine learning for Coronavirus Disease 2019 (COVID-19) diagnosis using chest computed tomography (CT) is crucial.
• Existing methods often lack high recognition accuracy and require extensive training data.
• Addressing these limitations is vital for practical clinical applications.

Purpose of the Study:

• To propose a novel COVID-19 recognition system using CT images.
• To achieve high recognition accuracy with limited training data.
• To enhance generalizability and stability in COVID-19 diagnosis.

Main Methods:

• A novel BCELoss function (LSFLLW-R) incorporating Label Smoothing, Focal Loss, and Label Weighting Regularization was developed to optimize solutions and prevent overfitting.
• A backbone network utilized two-phase contrastive self-supervised learning for multi-label classification.
• A decision-fusing ensemble learning method was employed to ensure system stability and balanced performance metrics.

Main Results:

• The system achieved 94.3% accuracy, 94.1% precision, 93.4% recall, 94.7% F1-score, and 98.9% AUC on the COVID-CT dataset.
• The proposed system effectively identifies pathological locations in CT images.
• Performance metrics demonstrate superior accuracy, generalizability, and stability compared to state-of-the-art methods.

Conclusions:

• The developed system offers a highly accurate and data-efficient approach for COVID-19 diagnosis via CT scans.
• The integration of advanced loss functions, self-supervised learning, and ensemble methods significantly improves diagnostic performance.
• This novel system presents a promising advancement for computer-aided COVID-19 diagnosis in clinical settings.