Image analysis model development.
The objective of this study was to develop a novel clinical workflow tool that uses model-based shape recognition technology to allow efficient, semiautomated detailed annotation of each vertebra between T4 and L4 on plain lateral radiographs.
Identification of prevalent vertebral fractures, especially when not symptomatic, has been problematic despite their importance. There is a recognized need to increase the opportunities to detect vertebral fractures so that clinically beneficial therapeutic interventions can be initiated.
Radiographs obtained from 165 subjects in the Canadian Multicenter Osteoporosis Study (CaMos) were used to construct a vertebral shape model of the vertebral column from T4 to L4 using a statistical learning technique, as well as to estimate the accuracy and precision of this automated software tool for vertebral shape analysis. Radiographs showing scoliosis greater than 15° were excluded.
Vertebral contours defined by 95 points per vertebra, represented by 79,895 points in total, were assessed on 841 individual vertebrae. The mean absolute accuracy error calculated over each vertebra in each test image was 1.06 ± 1.2 mm. This value corresponded to an average 3.4% of vertebral height. The mean precision error, reflecting interobserver variability, per vertebra of the resulting annotations was 0.61 ± 0.73 mm. This value corresponded to an average 2.3% of vertebral height. Accuracy and precision error estimates did not differ notably by vertebral level.
The results of the current study indicate that statistical modeling can provide a robust tool for the accurate and precise semiautomated annotation of vertebral body shape from T4 to L4 in patients who do not have scoliosis greater than 15°. This method may prove useful as a clinical workflow tool to aid the physician in vertebral fracture assessment and might contribute to decision-making about pharmacologic treatment of osteoporosis.
Using radiographs from 165 subjects in the Canadian Multicenter Osteoporosis Study, a vertebral shape model was constructed, and accuracy and precision of vertebral shape analysis were estimated. The mean accuracy and precision errors were 1.06 ± 1.2 mm and 0.61 ± 0.73 mm, respectively. Statistical modeling may provide a robust tool for the automated annotation of vertebral body shape.
From the *Optasia Medical Ltd, Cheadle, Cheshire, United Kingdom; †Bio-Imaging Technologies Inc, Newton, PA; ‡Department of Radiology, VCU Medical Center, Medical College of Virginia, Richmond, VA; §Roche Pharmaceuticals, Nutley, NJ; ¶Novartis Pharmaceuticals, East Hanover, NJ; ∥Jon E. Block, PhD, Inc., San Francisco, CA; and **Department of Radiology, VU University Medical Center, Amsterdam, The Netherlands.
Acknowledgment date: December 9, 2008. First revision date: March 4, 2009. Second revision date: March 31, 2009. Acceptance date: April 1, 2009.
The device(s)/drug(s) that is/are the subject of this manuscript is/are not FDA-approved for this indication and is/are not commercially available in the United States.
Corporate/Industry funds were received in support of this work. One or more of the author(s) has/have received or will receive benefits for personal or professional use from a commercial party related directly or indirectly to the subject of this manuscript: e.g., honoraria, gifts, consultancies.
Supported by Optasia Medical (Burlington, MA), Novartis Pharmaceuticals (East Hanover, NJ), and Bio-Imaging Technologies Inc (Newton, PA).
Address correspondence and reprint requests to Jon E. Block, PhD, 2210 Jackson St, 401, San Francisco, CA 94115; E-mail: email@example.com