Classification of vertebral compression fractures using CT-based radiomics: a viable alternative to MRI

Abstract Objects To evaluate whether CT-based machine learning models can accurately detect and differentiate fresh from old vertebral compression fractures (VCFs). Methods From June 2018 to December 2024, patients with VCFs from two hospitals who underwent both CT and MRI within one week of each other were retrospectively enrolled. Vertebrae were automatically segmented, and radiomics features were extracted. Two models were developed: one for VCF detection and another for classifying fresh versus old fractures. The classification model was trained with two approaches and producing three-level probability outputs. Subgroup analyses by compression grade and shape were performed. Clinical applicability was assessed, and performance was evaluated using receiver operating characteristic (ROC) analysis. Results This study analyzed data from 716 patients (mean age, 64.9 ± 9.0 years; 269 males) involving 2,852 vertebrae. In VCF detection, the machine learning model performed strongly, achieving an AUC of 0.944 in the external test set. With the inclusion of an uncertainty category, the VCF classification model achieved its best performance, with an AUC of 0.892 in the external test set. Subgroup analyses showed the highest performance for grade 2 fractures (AUC, 0.891) and crush-type fractures (AUC, 0.911) in the external test set. Moreover, AI assistance significantly improved radiologists’ diagnostic accuracy, particularly for junior radiologist, increasing the AUC from 0.628 to 0.793 ( p < 0.001). Conclusion The developed models not only achieved high accuracy in detecting vertebral compression fractures and differentiating fresh from old lesions, but also enhanced human–machine collaboration, thereby assisting junior radiologists in narrowing the diagnostic gap with senior radiologists.