Clinical Scorecard: Assessment of Lung Involvement in COVID-19 Pneumonia Using a Dual U-Net Approach: Training and Evaluation Across Various Datasets with Distinct Annotation Standards

At a Glance

Key Highlights

• COVID-19 lung abnormalities include bilateral ground-glass opacifications, consolidations, crazy-paving patterns, reversed halo sign, and vascular enlargement.
• The LungQuant system uses a cascade of two U-Nets: the first segments lungs, the second segments COVID-19 lesions, providing quantitative lung involvement and CT severity score (CT-SS).
• Training on heterogeneous, publicly available datasets with different annotation standards improves generalization of the DL segmentation model.

Guideline-Based Recommendations

Diagnosis

• Use chest CT imaging to identify COVID-19 related lung abnormalities including GGO and consolidations.
• Apply automated DL-based segmentation tools like LungQuant to quantify lung involvement objectively.

Management

• Utilize quantitative lung lesion volume and CT-SS from segmentation outputs to assist clinical decision-making and disease severity assessment.

Monitoring & Follow-up

• Monitor changes in lung lesion volume and CT-SS over time using serial CT scans and DL segmentation to evaluate disease progression or resolution.

Risks

• Variability in CT acquisition protocols and annotation standards can affect segmentation accuracy.
• Data conversion processes (e.g., DICOM to NifTI) may reduce image resolution and impact quantitative analysis.

Patient & Prescribing Data

Patients with confirmed or suspected COVID-19 undergoing chest CT

Quantitative assessment of lung involvement via DL segmentation may guide treatment intensity and monitor response but requires validation in clinical workflows.

Clinical Best Practices

• Train DL segmentation models on large, heterogeneous datasets to improve robustness and generalizability.
• Use dual-stage U-Net architecture: first for lung segmentation, second for lesion segmentation within lung regions.
• Incorporate morphology-based refinement post lung segmentation to improve mask accuracy.
• Calculate CT severity score based on percentage of lung volume affected by lesions to standardize severity assessment.

References

• Lessmann et al. U-net model for lesion segmentation
• Fang et al. Automated lung segmentation and lesion detection
• Ma et al. Training DL systems on limited COVID-19 CT data
• Morozov et al. MosMed dataset description
• Ronneberger et al. U-net architecture