Radboudumc

Radiology departments at Radboudumc (Nijmegen) and Jeroen Bosch Hospital ('s-Hertogenbosch) face increasing workloads, with radiologists spending substantial time reporting chest radiographs that turn out to be normal. There was a need to identify whether AI could reliably triage normal studies to reduce unnecessary reporting burden without missing critical findings.

Researchers retrospectively evaluated the commercially available Lunit INSIGHT CXR3 AI system across 1,670 consecutive chest radiographs from both hospitals. The system was assessed for its ability to correctly identify normal radiographs using ROC analysis, and its sensitivity for urgent and critical findings was also measured to ensure patient safety under any triage workflow.

The AI achieved an AUC of 0.918 for detecting normal chest radiographs, with no significant performance difference between the two hospital sites. At a conservative threshold, the system identified 53% of normal radiographs, translating to a potential 15% reduction in radiologist reporting workload. The system also demonstrated a negative predictive value (NPV) of 98% for urgent and critical findings, indicating no critical cases would be missed.

• A single commercially available AI system can generalize well across multiple hospital sites with consistent performance, supporting multi-center deployment.
• Conservative operating thresholds can capture meaningful workload savings (15%) while preserving near-complete safety for critical cases (98% NPV).
• Regulatory approval for autonomous use remains a barrier; radiologist supervision is still required, limiting the full potential workload reduction until clearance is obtained.