Bulent Yildiz, Emre Delibaş, Irfan Atik, Omer Abdulaziz Bilgec
Background/Objectives: Osteoporosis and osteopenia remain underdiagnosed despite routine CT examinations containing potentially useful bone information. This study evaluated the feasibility of patient-level three-class classification of Normal, Osteopenia, and Osteoporosis classes using lumbar vertebral PNG images derived from routine thoracoabdominal CT scans, without raw DICOM data, HU values, or quantitative BMD information. Methods: The dataset comprised 190 patients labeled by DXA-derived mean L1-L4 T-scores and 760 vertebral images. Four vertebral images per patient were analyzed jointly using patient-level 5-fold stratified cross-validation. ROI annotations defined only the standardized crop center and were not used as model inputs, masks, or features. ResNet18, ResNet50, and DenseNet121 transfer learning models were trained with patient-level L1-L4 feature fusion, and their probabilities were combined using equal-weight soft voting. Results: Within the patient-level 5-fold cross-validation framework (the same fold was used for model selection and final performance reporting, rather than serving as an independent held-out test set), the final ensemble reached a cross-validation accuracy of 0.7474 (95% CI: 0.6842-0.8053), balanced accuracy of 0.7377 (95% CI: 0.6703-0.8009), macro-F1 of 0.7403 (95% CI: 0.6730-0.8011), and macro-average ROC-AUC of 0.8576 (95% CI: 0.8091-0.9013). ROI occlusion reduced performance, whereas mask-only and geometry-only controls indicated that ROI annotation or geometry alone was insufficient for class discrimination. Grad-CAM showed that activations were not confined to ROI annotations, although residual ROI-related effects could not be fully excluded. Conclusions: Patient-level multi-vertebral transfer learning demonstrated moderate feasibility for image-based osteoporosis decision support, supporting a preclinical decision-support role rather than direct diagnosis or replacement of QCT/BMD assessment.