An automated algorithm for quantitative morphometry of thoracic and lumbar vertebral bodies in lateral radiographs.

This exploratory study developed and evaluated an artificial intelligence (AI)-based algorithm for quantitative morphometry to assess vertebral body deformities indicative of fractures. To achieve this, 709 radiographs from 355 cases were utilized for algorithm development and performance evaluation. The proposed algorithm integrates a first-stage AI model to identify the positions of thoracic and lumber vertebral bodies in lateral radiographs and a second-stage AI model to annotate 6 landmarks for calculating vertebral body height ratios (, , and ). The first-stage AI model achieved a sensitivity of 97.6%, a precision of 95.1%, and an average false-positive ratio of 0.43 per image for vertebral body detection. In the second stage, the algorithm's performance was evaluated using an independent dataset of vertebrae annotated by 2 spine surgeons and 1 radiologist. The average landmark errors ranged from 2.9% to 3.3% on the X-axis and 2.9% to 4.0% on the Y-axis, with errors increasing in more severely collapsed vertebrae, particularly at central landmarks. Spearman's correlation coefficients were 0.519-0.589 for , 0.558-0.647 for , and 0.735-0.770 for , comparable with correlations observed among human evaluators. Bland-Altman analysis revealed systematic bias in some cases, indicating that the algorithm underestimated anterior and central height collapse in deformed vertebrae. However, the mean differences and limits of agreement between the algorithm and external evaluators were similar to those among the evaluators. Additionally, the algorithm processed each image within 10 s. These findings suggest that the algorithm performs comparably with human evaluators, demonstrating sufficient accuracy for clinical use. The proposed approach has the potential to enhance patient care by being widely adopted in clinical settings.