DRR4Covid

DRR4Covid generates infection-aware digitally reconstructed radiographs from labeled CT scans and trains a segmentation network with domain adaptation to segment and quantify COVID-19 infection in chest X-rays.

Key Features:

• Infection-aware DRR generator: Produces DRRs from labeled CT scans with pixel-level annotations of infected regions for training segmentation models.
• Segmentation network: Trains a network to learn and identify COVID-19 infection patterns from generated DRRs for CXR segmentation.
• Domain adaptation module: Applies domain adaptation to transfer and generalize the segmentation model from synthetic DRRs to real-world chest X-rays.
• Performance metrics: Reports classification Accuracy 0.949, AUC 0.987, F1-score 0.947 and segmentation Accuracy 0.956, AUC 0.980, F1-score 0.955 on a test set of 558 normal and 558 positive cases.
• Detection limit estimation: Estimates X-ray detection limits by adjusting radiological signs in infection-aware DRRs, reporting detectable infected lung volume ≈ 19.43% ± 16.29% and a lower bound of 20.0% infected voxel contribution rate for significant radiological signs.

Scientific Applications:

• Automated infection segmentation: Segments and quantifies COVID-19–related infected regions in chest X-rays without requiring CXR annotations.
• Quantitative performance evaluation: Enables quantitative evaluation of classification and segmentation performance on balanced CXR test cohorts.
• X-ray detection-limit assessment: Provides an empirical estimate of the minimum infected lung volume detectable by radiological signs in X-rays.

Methodology:

Train the segmentation network on infection-aware DRRs derived from labeled CT scans with pixel-level annotations. Apply a domain adaptation module to bridge the gap between DRRs and real CXRs. Adjust radiological signs in DRRs to estimate detection limits and evaluate performance with statistical analyses (p < 0.05).