Abstract / Summary
Objectives: The aim of this study is to provide early diagnosis of osteoporosis, osteopenia, and normal bone mineral density (BMD) on CT scans utilizing machine learning and radiomics techniques. Methods: In this retrospective study, 874 patients who met the predefined eligibility criteria were included. BMD was categorized as normal, osteopenia, or osteoporosis according to DEXA-derived T-scores. Lumbar vertebral bodies were automatically segmented on abdominal and lumbar CT images, and radiomic features were extracted from the segmented regions. Logistic regression (LR), decision table (DT), random forest (RF), multilayer perceptron (MLP), and bagging models were developed to distinguish among the three BMD categories. Model performance was evaluated using accuracy, F-score, area under the receiver operating characteristic curve (ROC area), and area under the precision–recall curve (PRC area). Results: Notably, the best results were observed in the L3 vertebra among the 874 patients included in the study. Accordingly, the MLP and LR methods showed the greatest accuracy value of 83.5%, F-score value of 81.2%, ROC area value of 0.912, and PRC area value of 0.892 in cases with normal BMD. At the L2 vertebral level, the LR approach produced the greatest results for patients with osteoporosis (Acc: 72.5%, F-score: 72.5%, ROC area: 0.897, and PRC area: 0.797). Conclusions: In the three-class BMD assessment, the models showed the highest performance for normal BMD and the lowest performance for osteopenia. Among the individual lumbar levels, L3 demonstrated the highest radiomics-based classification performance, whereas L1 showed comparatively lower performance. These level-specific findings should be regarded as exploratory and should not be interpreted as supporting clinical BMD classification from a single vertebra. CT radiomics may have potential as an opportunistic screening tool to identify patients who may benefit from formal DEXA or further osteoporosis evaluation; however, patient-level multivertebral validation is required before clinical implementation.