Audit of Artificial Intelligence Assisted Diagnosis of Vertebral Compression Fractures Using Dual-energy X-ray Absorptiometry

Purpose or Learning Objective: To determine the accuracy of artificial intelligence (AI)-assisted dual-energy X-ray absorptiometry (DEXA) vertebral fracture analysis.

Methods or Background: We used an AI-assisted assessment of 2,184 DEXA vertebral fracture assessments undertaken on UK Biobank volunteers (application no. 17295).

An AI algorithm was used to assess cases through an automated process of image initialization to detect the centers of the vertebral bodies, high-resolution fitting to define the outline and edges of the vertebra, and classification of diagnosis.

The AI system compared vertebral height measurements with Genant's definitions to classify mild, moderate, and severe fractures. An annotator then reviewed the output of the AI classification, determining whether the vertebra was fractured based on the overall appearance. This method permitted exclusion of the machine allowances for abnormalities not due to a fracture, for example a congenital abnormality or Schmorl's node, producing the final AI-assisted fracture classification.

A musculoskeletal radiologist with > 40 years of experience audited the results.

Results or Findings: A total of 2,184 individuals were examined (men: 47%; women: 53%; age 44–79 years).

Detection of Schmorl's nodes and congenital vertebral anomalies was accurately relabeled by the annotator in most cases. Performance varied between annotators in detecting spinal tilt.

The radiologist was able to fully visualize 97.67% of the vertebrae examined; the AI-assisted system was able to analyze 99.29%.

A discrepancy was found in the classification of fracture status (i.e., binary classification as fracture [any grade] versus no/other pathology) in 614 vertebrae (2.17%). The radiologist diagnosed 8.9% (1,474) fewer vertebrae as fractured than the AI-assisted system (1,618). In the classification of patients as having fractures (any grade) or no fractures, with the radiologist as the gold standard, the AI-assisted system achieved a sensitivity, specificity, and accuracy of 84.9%, 91.0%, and 89.0%, with a tendency to overcall fractures (true negative: 1,330; false negative: 109; false positive: 131; true positive: 614).

The radiologist observed that AI-assisted interpretation often drew attention to mild biconcave fractures that otherwise might have been overlooked. Vertebral tilt was a major factor in overcalling of fractures by the AI-assisted system.

Conclusion: An AI-assisted diagnosis of vertebral fracture analysis using DEXA is practical. The AI-assisted system tended to overcall fractures where there was lateral tilt in the vertebral column. Adjustments to the classifier may potentially overcome these difficulties.

Publication History

Article published online:
26 May 2023