An Explainable AI System for Automated COVID-19 Assessment and Lesion Categorization from CT-scans

AI can match or exceed expert performance on specific diagnostic tasks: The system achieved 90% sensitivity and 93.5% specificity for COVID-19 detection from CT scans, outperforming three expert radiologists on the same dataset of 162 scans, demonstrating that AI tools can be reliable diagnostic aids when properly developed.

Explainability is essential for clinical trust: The researchers demonstrated that their AI model's decision-making process was interpretable—it focused on clinically recognized COVID-19 lesion patterns (crazy paving, consolidation, and ground glass opacities)—which validates that the AI was learning medically meaningful features rather than spurious correlations.

Proper system design significantly impacts performance: Prior lung and lobe segmentation improved overall performance by over 20 percentage points, highlighting that responsible AI development requires careful architectural choices and that breaking complex tasks into logical components can substantially enhance reliability.

Validation against expert comparison is crucial: By directly comparing AI results against three radiologists' assessments rather than just reporting standalone metrics, the researchers provided meaningful context for clinicians evaluating whether to trust the system in practice.

Lesion categorization remains challenging even for AI: While COVID-19 detection reached high accuracy, lesion categorization achieved only 84% accuracy, illustrating that AI systems have varying reliability across different subtasks and that clinicians should understand these performance differences.