Deep Learning in Vertebral Fracture Detection: Systematic Review and Meta-analysis of Subject- vs. Vertebra-Level
Mohammad-Reza Hosseini-Siyanaki1, Babak Ahmadi2, Hakki Serdar Sagdic1
1Department of Radiology, University of Florida, Gainesville, Florida (M.H., H.S.S., A.R., S.M., A.R., A.D., K.R.P.).
Academic Radiology
|December 6, 2025
Summary
Deep learning algorithms show high accuracy for detecting vertebral fractures. Subject-level models excel at screening with high sensitivity, while vertebra-level models offer precise localization and high specificity for diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Deep learning (DL) algorithms are increasingly used for vertebral fracture detection.
- Evaluating DL performance requires disentangling subject-level and vertebra-level approaches.
- Understanding technical and methodological factors is crucial for clinical application.
Purpose of the Study:
- To conduct a context-aware evaluation of DL algorithms for vertebral fracture detection.
- To quantify the influence of technical and methodological factors on DL performance.
- To guide clinical use and standardized reporting of DL tools.
Main Methods:
- A PRISMA-compliant systematic review of five databases (searched to February 2025).
- Inclusion of English-language studies reporting accuracy metrics for vertebral fracture detection.
- Risk of bias assessment using QUADAS-AI and hierarchical summary ROC models for meta-analysis.
Main Results:
- 36 studies (96,956 patients) were included; pooled subject-level sensitivity/specificity was 84%/91% (AUC 0.94), and vertebra-level was 80%/97% (AUC 0.96).
- Subject-level models prioritized sensitivity for screening; vertebra-level models achieved higher specificity for precise localization.
- External validation decreased sensitivity but maintained specificity; radiographs favored subject-level, CT favored vertebra-level analysis. Patient selection bias was common (61% of studies).
Conclusions:
- DL models demonstrate high accuracy for vertebral fracture detection.
- Subject-level DL is suitable for screening/triage (high sensitivity); vertebra-level DL is better for diagnosis (high specificity, precise localization).
- Clinical application requires aligning model granularity with task and context due to performance variability across modalities and data sources.
