How Accurate Is AI Radiology Reporting?
Evidence from Published Clinical Studies

The central question in AI radiology

"How accurate is AI radiology reporting?" is one of the most searched questions in the field — and one of the least honestly answered in vendor marketing. Most claims cite cherry-picked detection studies on single pathologies. The better question is: how does AI CT reporting perform across the full clinical workflow, including structured report generation, on a realistic mix of emergency and routine studies?

This article summarizes two independent peer-reviewed studies that evaluated AI CT reporting in real clinical settings, explains what the numbers mean in practice, and describes what "accuracy" should actually mean when selecting an AI radiology partner.

Study 1: Emergency chest CT in a clinical setting (Polish Radiology, 2025)

The most rigorous evaluation of xAID's AI CT reporting to date was published in Polish Radiology (Pol Radiol, 2025). The study was a retrospective single-center evaluation conducted at an emergency radiology department, using 90 consecutive unenhanced chest CT scans.

Study design: The same 90 cases were read by radiologists with AI assistance (AI-assisted) and without (unaided). Pathologies were evaluated across 9 categories:

• Lung nodules
• Pulmonary opacifications
• Pneumothorax
• Pleural and pericardial effusion
• Pulmonary artery dilatation
• Coronary artery calcifications
• Aortic diameter
• Vertebral fractures
• Rib fractures

Three specific categories showed the largest gains: AI outperformed unaided radiologists substantially in detecting coronary artery calcifications, pulmonary artery dilatation, and vertebral fractures — findings that are commonly missed on routine emergency reads because they require precise quantitative measurement rather than pattern recognition.

What this means for practice: A 33.9-percentage-point sensitivity advantage is clinically significant. In a center reading 200 chest CTs per month, this translates to approximately 68 additional correctly identified pathologies monthly — many of which would otherwise require patient recall, repeat imaging, or represent missed diagnoses.

Study 2: Multi-center European clinical utility assessment (ResearchGate, 2025)

A second independent study evaluated AI CT reporting clinical utility across four European radiology centers — France, Greece, Slovakia, and the United Kingdom. The study assessed 81 non-contrast chest CT cases with four board-certified radiologists.

Unlike the Polish Radiology study (which measured detection accuracy), this study focused on clinical integration: would radiologists actually trust and use AI-generated report elements in practice?

94.9% approval does not mean radiologists agreed with everything — it means they found AI-generated structured report elements clinically usable with minor modifications. 81.5% reported that AI segmentation data contributed positively to their diagnostic process.

The foundation model: technical architecture

The accuracy results above are driven by xAID's Swin transformer-based foundation model — a transformer architecture adapted for 3D volumetric medical imaging. Key technical specifications:

• Input resolution: Up to 256³ voxels — capturing fine-grained anatomical detail that lower-resolution models miss
• Coverage: Head, chest, and abdomen CT
• Findings analyzed: 100+ per scan
• Macro F1 score: 0.86 across clinically relevant pathologies
• Secondary verification: A second AI layer independently reviews findings; divergences are flagged for radiologist attention

What "accuracy" should mean when evaluating AI radiology providers

Most AI radiology vendors cite performance numbers from internal evaluations or narrow single-pathology studies. When evaluating accuracy claims, ask:

1. Is the study independent? Was it conducted by the vendor or by an independent research team? The Polish Radiology study and the European multi-center study were both independent.
2. What was the study design? Real emergency cases on a consecutive series are harder to cherry-pick than curated test sets.
3. What pathologies were evaluated? Single-pathology performance (e.g., nodule detection) does not predict full-scan reporting performance.
4. Is there published peer-reviewed evidence? Published performance numbers from independent studies — not vendor self-assessments — are required for a meaningful accuracy claim.
5. Is a radiologist reviewing every report? Autonomous AI reporting without human review introduces liability and accuracy risks not present in AI-assisted reporting.

Frequently asked questions