Bone density measurement in patients with spinal metastatic tumors using chest quantitative CT deep learning model

Abstract

Objective: This study aims to develop a deep learning model using the 3DResUNet architecture to predict vertebral volumetric bone mineral density (vBMD) from Quantitative Computed Tomography (QCT) scans in patients with spinal metastatic tumors, enhancing osteoporosis screening capabilities. Methods: 749 patients with spinal metastatic tumors underwent QCT vertebral vBMD measurements. The dataset was randomly split into training (599 cases) and test sets (150 cases). The 3DResUNet model was trained for vBMD classification and prediction using QCT images processed with automated bone segmentation and ROI extraction. Results: The deep learning model demonstrated strong performance with Spearman correlation coefficients of 0.923 (training set) and 0.918 (test set) between predicted and QCT-measured vBMD values. Bland-Altman analysis showed a slight bias of −1.42 mg/cm3 (training set) and −1.14 mg/cm3 (test set) between the model predictions and QCT measurements. The model achieved an area under the curve (AUC) of 0.977 (training set) and 0.966 (test set) for diagnosing Osteoporosis based on vBMD. Conclusion: The developed deep learning model using 3DResUNet effectively predicts vertebral vBMD from QCT scans in patients with spinal metastatic tumors. It provides accurate and automated vBMD measurements, potentially facilitating widespread osteoporosis screening in clinical practice, mainly where DXA availability is limited.

Keywords

• Quantitative computed tomography
• Spinal metastatic tumors
• Deep learning
• Deep red
• Volumetric bone mineral density
• Bone mineral content