A machine-learning radiomics model derived from lumbar spine computed tomography outperformed vertebral bone quality scoring and Hounsfield unit measurement in detecting osteoporosis.
In a retrospective study, researchers evaluated 166 patients who underwent concurrent dual-energy X-ray absorptiometry, lumbar spine computed tomography (CT), and magnetic resonance imaging within 30 days between January 2020 and June 2023. Osteoporosis was defined as a lumbar spine T-score of −2.5 or lower on dual-energy X-ray absorptiometry, according to World Health Organization criteria.
Among 664 initially assessed L1 to L4 vertebrae, 656 were included in the final analysis following the exclusion of inadequate regions of interest,. The researchers identified 210 (32%) vertebrae classified as osteoporotic. Vertebrae were allocated at the patient level to training (80%) and test (20%) cohorts. There were no statistically significant differences in osteoporosis prevalence between the cohorts (34% vs 26%).
Three-dimensional segmentation of vertebral bodies on CT enabled the extraction of 851 radiomic features per region of interest. Features demonstrating intraclass correlation coefficients greater than 0.90 for both interobserver and intraobserver reproducibility were retained, resulting in 450 stable features. Minimum redundancy–maximum relevance filtering reduced the feature set to 30 candidates, and least absolute shrinkage and selection operator regression identified nine features with nonzero coefficients to construct the radiomic signature.
Four supervised machine-learning classifiers were developed: logistic regression, support vector machine, extreme gradient boosting (XGBoost), and random forest. In receiver operating characteristic analysis, the XGBoost model demonstrated the highest performance, with an area under the curve of 0.89 in the training cohort and 0.91 in the test cohort. In the test cohort, sensitivity was about 89%, specificity was 82%, and the negative predictive value was 94%.
By comparison, models based on vertebral bone quality scores and Hounsfield unit measurements showed lower discriminative performance in both cohorts. DeLong testing indicated that the radiomics XGBoost model was more effective compared with vertebral bone quality and Hounsfield unit models in predicting osteoporosis. Decision curve analysis demonstrated greater net clinical benefit across probability thresholds for the radiomics model relative to the alternative approaches.
The selected radiomic features included first-order statistical measures and texture-based metrics derived from gray level co-occurrence, gray level dependence, gray level run length, gray level size zone, and neighborhood gray tone difference matrices, reflecting multidimensional characteristics of vertebral bone marrow microstructure.
The researchers highlighted that the proposed model is intended to be used as a screening tool rather than a replacement for dual-energy X-ray absorptiometry. Limitations included the retrospective, single-center design; modest sample size; use of a single CT scanner and acquisition protocol; and inability to stratify patients with osteopenia. External validation in multicenter cohorts was identified as necessary to confirm generalizability.
“The radiomics-XGBoost model demonstrated superior predictive accuracy,” noted lead study author Cheng Gao, of the Department of Orthopedics at The Affiliated Jiangning Hospital with Nanjing Medical University in China, and colleagues.
The authors reported no commercial or financial relationships that could be viewed as potential conflicts of interest, and Jinhui Liu, of the Nanjing Medical University, disclosed a shared parent institutional affiliation with the study authors during the review process.
Source: Frontiers in Medicine
