An End-to-End Radiomic Framework for Automatic Vertebral Lesion Classification and 3D Visualization

Early and reliable identification of vertebral metastases on computed tomography remains a major challenge in oncologic imaging due to the morphological complexity of metastatic lesions and the high inter-patient variability of spinal anatomy. In this study, an end-to-end interpretable radiomic-based framework was developed to automatically distinguish healthy from metastatic vertebrae using segmented DICOM data, coupled with an interactive virtual reality (VR) visualization module implemented in Unity 3D. The proposed framework integrates radiomic feature extraction and selection, informed undersampling to address class imbalance, and automatic machine learning-based classification. To facilitate interpretation, patient-specific 3D models with overlapped classifier outputs were integrated into a VR desktop application, enabling advanced exploration of patient-specific spinal models, with color-coded visualization of algorithmic predictions and expert-defined suspicious lesions. The final classification model, trained using a Random Forest algorithm and optimized via stratified 5-fold cross-validation, achieved an overall accuracy of 0.86, an Area Under the Receiver Operating Characteristic Curve of 0.91, and an F1-score of 0.81 for the metastatic class on the independent test set, achieving competitive diagnostic performance while preserving transparency and clinical interpretability. This study represents a foundational step toward intelligent, interactive, and clinically interpretable tools for the diagnosis and follow-up of spinal metastatic disease.