关键词: 布鲁氏菌性脊柱炎, 化脓性脊柱炎, 影像组学, 机器学习, 联合模型

Abstract: Objective To evaluate the diagnostic value of Logistic regression (LR) and support vector machines (SVM) models, constructed based on radiomic features from CT and MRI T₂-weighted imaging (T₂WI) combined with clinical features in the differential diagnosis between brucella spondylitis (BS) and pyogenic spondylitis (PS). Method Patients diagnosed with BS or PS by etiology or pathology who underwent both CT and MRI T₂WI examinations at Inner Mongolia Autonomous Region People's Hospital from January 2017 to August 2025 were retrospectively collected. Patients were randomly divided into a training set (n=70, 37 BS cases, 33 PS cases) and a test set (n=30, 16 BS cases, 14 PS cases) at a ratio of 7:3. Univariate and multivariate LR were used to screen clinical risk factors, and a clinical model was established accordingly. The involved vertebral bodies were delineated on CT and MRI T₂WI images for radiomic features extraction. Key features were selected from the training set by statistical testing, least absolute shrinkage and selection operator regression, and cross-validation to calculate radiomic scores (radscore, RS), thus establishing a dual-modal radiomic model of CT-MRI. Clinical features were fused with RS, and combined clinical-radiomic models were established using LR and SVM, respectively. Receiver operating characteristic curves of each model were plotted, and the diagnostic efficacy was evaluated by indicators such as the area under the curve (AUC). The Delong test was used to compare the differences in AUC values among the models. Result bony destruction of spinal appendages and paraspinal abscess were key clinical features for differentiating BS from PS. The clinical feature model showed AUCs of 0.718 (training set) and 0.612 (test set). A total of 1409 radiomic features were extracted from CT and MRI T₂WI images respectively; after feature selection, 10 key CT features and 5 key MRI T₂WI features were retained, and the radiomic model achieved AUCs of 0.897 (training set) and 0.732 (test set). For the combined clinical-radiomic models, the LR model had AUCs of 0.901 (training set) and 0.781 (test set), while the SVM model had AUCs of 0.902 (training set) and 0.777 (test set). Delong test indicated no statistically significant differences in diagnostic efficacy between the LR model and SVM model(P>0.05). Conclusion LR and SVM models based on CT and MRI T₂WI radiomics combined with clinical features show favorable differential diagnostic performance for BS and PS, and can provide a non-invasive reference for clinical differentiation.

Key words: Brucella spondylitis, Pyogenic spondylitis, Radiomics, Machine learning, Combined model