Commercial AI for Opportunistic Detection of Vertebral Compression Fractures on Routine CT: A Systematic Review of Diagnostic Performance and Clinical Yield

Technical Summary: Commercial AI for Opportunistic Detection of Vertebral Compression Fractures on Routine CT

Problem Statement
Vertebral compression fractures (VCFs) are the most common osteoporotic fractures globally, carrying significant risks of increased mortality and subsequent fracture. Despite this, routine Computed Tomography (CT) scans performed for other clinical indications (e.g., pulmonary, abdominal, or oncological) frequently fail to report moderate-to-severe VCFs. Literature indicates that 38–63% of such fractures go undetected because radiologist attention is directed toward the primary clinical indication rather than the spine. While commercially available Artificial Intelligence (AI) software has emerged to address this "detection gap" via opportunistic screening, evidence regarding its diagnostic performance and real-world yield remains fragmented across different products, populations, and study settings. There is a lack of synthesized data on whether these regulatory-cleared tools reliably detect incidental fractures and whether their deployment translates into clinical action.

Methodology
This study is a systematic review conducted in accordance with PRISMA 2020 guidelines and registered with PROSPERO (CRD420261423256).

• Search Strategy: A comprehensive search was performed across PubMed, Scopus, and Web of Science (January 1, 2020 – May 31, 2026) for original peer-reviewed publications.
• Inclusion Criteria: Studies must evaluate commercially available, regulatory-cleared (FDA/CE) AI software for VCF detection on routine CT in adult patients. The studies required a valid reference standard (expert radiologist consensus or standardized grading) and quantitative diagnostic metrics.
• Exclusion Criteria: Non-commercial algorithms, research-only tools, MRI/X-ray/DXA modalities, conference abstracts without data, and studies lacking diagnostic accuracy metrics were excluded.
• Data Synthesis: Eight studies met the inclusion criteria, encompassing 6,182 patients or CT scans. Due to substantial heterogeneity in study design, patient populations, AI platforms, and outcome metrics, a formal meta-analysis was not performed. Instead, a structured narrative synthesis was conducted.
• Bias Assessment: Risk of bias was evaluated using the QUADAS-2 tool, focusing on patient selection, index test, reference standard, and flow/timing.

Key Contributions

• Platform Evaluation: The review evaluates four specific commercial AI platforms: Nanox HealthVCF (formerly Zebra HealthVCF), IB Lab FLAMINGO, Avicenna.AI CINA-VCF, and Nanox HealthOST.
• Quantification of the Detection Gap: The review systematically quantifies the discrepancy between routine radiology reports and AI detection, confirming that clinical reports miss a substantial proportion of fractures.
• Performance Stratification: The analysis differentiates performance based on fracture severity (Genant grades), detection thresholds, and the unit of analysis (patient-level vs. scan-level vs. vertebra-level).
• Evidence Gap Identification: The review highlights that while diagnostic detection performance is documented, no included study evaluated patient outcomes (e.g., treatment initiation, mortality) as a primary endpoint, limiting claims regarding clinical benefit.

Results

• Diagnostic Performance:
  • Sensitivity: Patient-level sensitivity for clinically actionable fractures (Genant grade ≥ 2) ranged from 65% to 95% across platforms. Performance was generally higher for severe deformities (Genant grade 3) and more variable for moderate fractures.
  • Specificity and NPV: Specificity was consistently high (87%–99%), and Negative Predictive Values (NPV) ranged from 89% to 98%, suggesting a low rate of missed fractures when the AI returns a negative result.
  • PPV Variability: Positive Predictive Values (PPV) varied significantly (44%–98%), primarily driven by cohort-specific fracture prevalence rather than algorithm performance.
• Radiologist Miss Rates: In studies comparing AI to routine reports, clinical radiology reports missed between one-third and two-thirds of moderate-to-severe fractures. AI consistently identified the majority of these missed cases. For example, in one study, routine reports documented only 36.7% of confirmed fractures, while AI recovered 88.2% of the undetected cases.
• Severity Dependence: Sensitivity improved as fracture severity increased. For instance, one study showed sensitivity rising from 89.7% for Grade 2 fractures to 99.0% for Grade 3 fractures.
• Risk of Bias: The "Index Test" domain was rated low risk across all studies. However, concerns were noted in "Patient Selection" (cohort enrichment), "Reference Standard" (non-independent verification in some studies), and "Flow and Timing" (retrospective designs, algorithm failure exclusions). One study (Behanova 2026) was rated high risk because its initial reference standard (routine reports) failed to identify true fractures that were later confirmed by expert re-evaluation.

Significance and Claims
The paper concludes that commercially available AI tools for opportunistic VCF detection on routine CT successfully identify fractures systematically missed in routine reports, demonstrating high specificity and negative predictive values. However, the authors emphasize several critical limitations:

1. Retrospective Nature: All included evaluations were retrospective; no study established whether AI deployment improves patient outcomes (e.g., treatment rates or mortality).
2. Moderate Fracture Variability: Performance is strongest for severe deformities but more variable for moderate fractures, which are the most clinically contested regarding treatment decisions.
3. Deployment Model: The evidence supports a "supervised screening aid" model where AI flags results for radiologist review prior to clinical action, rather than an autonomous diagnostic tool.
4. Regulatory Context: While the evaluated platforms held regulatory clearance at the time of study, the authors note that clearance status and approved indications must be verified locally before deployment.

The paper asserts that these tools may be integrated into structured osteoporosis care pathways to close the detection gap, but prospective outcome studies with standardized reporting are required before broad clinical adoption can be recommended.