Performance metrics from published validation studies. Real-world performance may differ. Educational reference, not product evaluation. Updated August 2026. See ACR AI Central for a comprehensive registry.
FDA-Cleared CXR AI Tools
| Product | Manufacturer | Pathologies | Mode | Training Risk |
|---|---|---|---|---|
| Annalise CXR | Annalise.ai | 124 findings (broadest) | Answer-delivery | High |
| Lunit INSIGHT CXR | Lunit Inc. | 10+ abnormalities | Answer-delivery + heatmap | High |
| qXR | Qure.ai | 29+ abnormalities | Answer-delivery + triage | High |
| VUNO Med | VUNO Inc. | 6+ major findings | Answer-delivery | Medium |
| BriefCase CXR | Aidoc | PTX triage, critical findings | Triage | Medium |
| Red Dot | Behold.ai | Normal vs. abnormal | Triage (binary) | Low |
Understanding Risk Levels
High = Finding-level diagnoses (answer-delivery). AI provides specific findings. Highest risk to independent skill development because the trainee's cognitive work is substantially replaced.
Medium = Triage or limited findings. AI handles detection/prioritization but the trainee still interprets. Moderate risk.
Low = Binary classification (normal vs. abnormal). The trainee must identify all specific findings. Lowest risk to independent skill development.
Performance by Pathology
| Pathology | Sensitivity | Specificity | Reliability | Training Risk |
|---|---|---|---|---|
| PTX (moderate-large) | 90–97% | 92–98% | Reliable | High |
| PTX (small/subtle) | 65–80% | 85–95% | Variable | High |
| Consolidation | 85–95% | 80–92% | Reliable | High |
| Pleural effusion | 88–96% | 90–96% | Reliable | Medium |
| Lung nodules | 70–90% | 75–90% | Variable | High |
| Cardiomegaly | 85–95% | 80–92% | Variable | Medium |
| Rib fractures | 60–85% | 75–90% | Variable | Medium |
| Mediastinal | 40–65% | Variable | Unreliable | Low |
| Bone lesions | 30–55% | Variable | Unreliable | Low |
Risk Mapping
Highest risk: pathologies where AI is good enough to replace search but not good enough to catch everything.
Highest Risk: AI is good but imperfect
Lung nodules, subtle pneumothorax, small consolidations. AI detects most (80–95%) so trainees stop searching independently, but it misses the subtle cases that require independent skill. The gap between AI's good-enough performance and 100% is where overreliance can undermine independent skill development; one form of this has been described as never-skilling.
Moderate Risk: AI creates miscalibration
Cardiomegaly, rib fractures, effusions. AI is generally good but has systematic biases (e.g., overcalling cardiomegaly on AP films, false-positive fractures at costochondral junctions). Primary risk: mis-skilling from absorbed errors.
Lower Risk: AI has limited coverage
Mediastinal abnormalities, bone lesions, soft tissue. Trainees must maintain independent assessment here. However, there is a risk of false completeness — believing a clean AI report means a normal study when AI simply does not evaluate these areas.
Key Questions About Your Department's AI
What pathologies does the AI detect?
Every AI tool has a defined detection list. Anything not on it is your responsibility alone. Know the boundaries of what your department's AI covers.
What is the false-negative rate for subtle findings?
Published sensitivity often reflects clear-cut cases. Subtle findings — small nodules, early architectural distortion, tiny pneumothoraces — may have significantly lower detection rates.
Triage mode or answer-delivery mode?
Triage preserves more of the trainee's diagnostic reasoning. Answer-delivery replaces more cognitive work. Know which mode your department's AI operates in — this shapes the training-risk profile.
Any known false-positive patterns?
Systematic false positives create mis-skilling risk. If the AI consistently overcalls certain findings, trainees may internalize those error patterns. Ask your attendings about known biases in your department's AI tools.