Authors: Jacob M. Mostert, Stephan Romeijn, Petra Dibbets-schneider, Daphne D. D. Rietbergen, Lenka M. Pereira Arias-Bouda, Christoph Götz, Matthew DiFranco, Hans Peter Dimai & Willem Grootjans
Abstract
Purpose
To investigate the time and effort needed to perform vertebral morphometry, as well as inter-observer agreement for identification of vertebral fractures on vertebral fracture assessment (VFA) images.
Methods
Ninety-six images were retrospectively selected, and three radiographers independently performed semi-automatic 6-point morphometry. Fractures were identified and graded using the Genant classification. The time needed to annotate each image was recorded, and reader fatigue was assessed using a modified Simulator Sickness Questionnaire (SSQ). Inter-observer agreement was assessed per-patient and per-vertebra for detecting fractures of all grades (grades 1–3) and for grade 2 and 3 fractures using the kappa statistic. Variability in measured vertebral height was evaluated using the intraclass correlation coefficient (ICC).
Results
The per-patient agreement was 0.59 for grades 1–3 fracture detection, and 0.65 for grades 2–3 only. The agreement for per-vertebra fracture classification was 0.92. Vertebral height measurements had an ICC of 0.96. The time needed to annotate VFA images ranged between 91 and 540 s, with a mean annotation time of 259 s. Mean SSQ scores were significantly lower at the start of a reading session (1.29; 95% CI: 0.81–1.77) compared to the end of a session (3.25; 95% CI: 2.60–3.90; p < 0.001).
Conclusion
Agreement for detection of patients with vertebral fractures was only moderate, and vertebral morphometry requires a substantial time investment. This indicates that there is a potential benefit for automating VFA, both in improving inter-observer agreement and in decreasing reading time and burden on readers.
Data availability
The dataset from this study is not publicly available due to data protection regulations.
Code availability
Statistical analyses were performed using open-source python modules Scikit-learn 0.20.3, Statsmodels 0.10.1, and Pingouin 0.3.8.
Download and read the full paper you can here