COVID-CT-Mask-Net

COVID-CT-Mask-Net predicts COVID-19 and distinguishes it from common pneumonia and control cases by performing instance segmentation and regional-feature-driven classification of CT scan lesions such as ground glass opacity and consolidation.

Key Features:

• Two-stage regional approach: Identifies and ranks lesion regions within CT images to inform classification.
• Mask R-CNN instance segmentation: Detects and segments ground glass opacity and consolidation, producing bounding boxes and masks.
• Ranked regional prediction-based classification: Uses ranked detections to classify images into three classes: COVID-19, common pneumonia, and control.
• Dataset: Trained and evaluated on the COVIDx-CT dataset from the China National Center for Bioinformation.
• Data efficiency: Reported strong generalization while training on approximately 5% of the available data.
• Performance metrics: Reported sensitivities of 90.80% (COVID-19), 91.62% (common pneumonia), 92.10% (normal) with 91.66% overall accuracy on a 21,182-image test set, and later evaluations reporting 93.88% COVID-19 sensitivity, 95.06% common pneumonia sensitivity, 96.91% true negative rate, and 95.64% overall accuracy on 21,192 CT scans.
• Segmentation accuracy: Best model achieved mean average precision of 44.68% using the MS COCO criterion across accuracy thresholds.
• Non-Maximum Suppression analysis: Explores the impact of NMS on both segmentation and classification accuracy to optimize handling of overlapping predictions.

Scientific Applications:

• COVID-19 detection from CT: Automated identification of COVID-19-consistent lesions in chest CT scans.
• Differential diagnosis: Distinguishes COVID-19 from common pneumonia and normal (control) cases.
• Lesion-level analysis: Segments ground glass opacity and consolidation for region-level diagnostic assessment.
• Method evaluation: Assesses effects of lesion mask handling and NMS on segmentation and classification performance.

Methodology:

Employs a two-stage process where Mask R-CNN performs instance segmentation to detect and rank bounding boxes and masks of ground glass opacity and consolidation, and classification is performed using ranked regional predictions; experiments include merging, excluding, or separately treating lesion masks, evaluation with the MS COCO mean average precision criterion, and analysis of Non-Maximum Suppression effects.