A Deep Learning Model Based on CT Images for Differentiating Acute and Chronic Osteoporotic Vertebral Compression Fractures

This study is a retrospective, multicenter observational study designed to develop and evaluate a deep learning-based system for differentiating acute and chronic osteoporotic vertebral compression fractures using computed tomography (CT) images.

Patients diagnosed with osteoporotic vertebral compression fractures who underwent both CT and magnetic resonance imaging (MRI) examinations will be retrospectively collected from the First Affiliated Hospital of Chongqing Medical University and one additional medical center between January 2023 and September 2025. Clinical data, including age, sex, and dual-energy X-ray absorptiometry (DXA) results, as well as complete DICOM-format CT and MRI images, will be collected. The interval between CT and MRI examinations must be less than two weeks. Patients with pathological fractures caused by infection or tumor, the presence of foreign materials such as bone cement or metallic hardware, or poor image quality with significant artifacts will be excluded.

The study workflow includes data collection, model development, performance evaluation, and model interpretability analysis. Multiple deep learning segmentation models, including U-Net, U-Mamba, and UNETR++, will first be evaluated for vertebral body segmentation performance. Based on the optimal segmentation results, classification models such as VGG-16, DenseNet-121, Vision Transformer (ViT), and Transformer-based architectures will be trained to differentiate acute and chronic compression fractures. The best-performing model will be selected to construct the final classification system.

Model performance for segmentation tasks will be assessed using Dice similarity coefficient and loss values. Classification performance will be evaluated in an external validation dataset using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value. Receiver operating characteristic curves and confusion matrices will be generated to visualize model performance.

To improve model interpretability, gradient-weighted class activation mapping (Grad-CAM) will be applied to generate heatmaps highlighting image regions that contribute most to model predictions. These heatmaps will be overlaid on CT images to visually demonstrate how the model differentiates acute and chronic osteoporotic vertebral compression fractures.