Automated stage classification of vertebral compression fractures using multimodal X-ray image analysis
Shenyang Duan1, Yufeng Deng1, Qingjiang Pang2
1College of Science and Technology, Ningbo University, Ningbo, China.
Objective:
To investigate a multimodal prediction model (DLRCM) based on X-ray images combined with deep learning, radiomics, and clinical data for identifying the stage of vertebral compression fractures (VCFs).
Method:
This study included X-ray images from 2,129 patients with vertebral compression fractures (VCFs) at Ningbo Second Hospital. (Ethics approval No.: SL-NBEY-KY-2025-254-01, from the Ethics Committee of Ningbo Second Hospital). According to MRI examinations, there were a total of 2,787 vertebrae with compression fractures. We allocated the training set, validation set, and test set in an 8:1:1 ratio to evaluate their performance. We used pre-trained ResNet34, ResNet50, VGG16, and VGG19 models to extract deep learning features of the vertebrae, and PyRadiomics to extract radiomics features. Clinical data features were obtained from the hospital. The extracted deep learning and radiomics features were fused with the patients' clinical data features. The fused features were then used for prediction using five machine learning methods: logistic regression (LR), random forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) for prediction, and compared the experimental performance of different combinations.
Results:
Through experimental comparison, the best experimental results were obtained by using VGG16 to extract deep learning features, then integrating radiomics and clinical data features, and finally performing regression prediction using LR. The accuracy rate reached 0.986, and the AUC reached 0.988 (95% CI: 0.975-0.990). Additionally, decision curve analysis (DCA) further indicated that the model provides high net clinical benefit across most probability threshold ranges. These results suggest that the model we constructed has good clinical applicability and interpretability, and can provide auxiliary diagnostic evidence for orthopaedic or radiology physicians in assessing the stage of vertebral compression fractures (VCFs), thereby guiding the formulation of subsequent treatment strategies.
Conclusion:
This study validated the effectiveness and practical value of our method in identifying the stage of vertebral compression fractures (VCFs) in X-ray images. The results showed that the method has high predictive performance and is expected to provide auxiliary decision support for identifying the stage of vertebral compression fractures (VCFs), promoting timely intervention and the formulation of targeted treatment strategies.
