MRI Radiomics and Deep Learning for Differentiating Spinal Tuberculosis from Spinal Metastasis: A Multicenter Study.
Authors
Affiliations (5)
Affiliations (5)
- Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China.
- Shanghai Changzheng Hospital, Naval Medical University, Shanghai, China.
- Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China. [email protected].
- Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China. [email protected].
- Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China. [email protected].
Abstract
To develop radiomics model, 2D deep learning (2D DL) model, 2.5D deep learning (2.5D DL) model, 3D deep learning (3D DL) model, and their fusion models to differentiate spinal tuberculosis (STB) from spinal metastasis (SM) based on MR imaging. This retrospective study included 186 STB patients and 219 SM patients. Patients from three hospitals were divided into a training set (n = 191), an internal test set (n = 82) and an external test set (n = 132). Using the ResNet50 network pre-trained leveraging the ImageNet dataset, we developed 2D, 2.5D, and 3D deep learning models based on MR imaging. We constructed the radiomics model by using PyRadiomics to extract radiomics features. Deep learning models and radiomics model were used to construct two fusion models: the feature-based model (Rad_DL_FB) and the decision-based model (Rad_DL_DB). And the accuracy of these models was validated. In the internal test set, the Rad_DL_FB model demonstrated superior performance compared to other models, achieving the highest AUC of 0.971 (95% CI: 0.953-0.989) and accuracy of 0.927. In the external test set, the Rad_DL_DB model demonstrated superior performance compared to other models, achieving the highest AUC of 0.961 (95% CI: 0.940-0.982) and accuracy of 0.909. Furthermore, fusion-model assistance significantly improved the diagnostic accuracy of junior spine surgeons and narrowed the observed performance gap between junior and senior surgeons. The fusion models including Rad_DL_FB model and Rad_DL_DB model, which integrates radiomics features and deep learning features of different dimensions, can effectively differentiate STB from SM, thereby assisting clinicians in making clinical decisions.