Differentiation model for spinal tuberculosis and pyogenic spondylitis based on quantitative SPECT/CT and multimodal machine learning

To develop and evaluate a multimodal machine learning model integrating clinical, laboratory, conventional imaging, and quantitative SPECT/CT-derived parameters for differentiating spinal tuberculosis (STB) from pyogenic spondylitis (PS). This retrospective study included 96 patients with STB and 69 with PS who underwent quantitative 99m Tc-MDP SPECT/CT between September 2024 and May 2026. Demographic data, laboratory indicators, conventional imaging features, and quantitative SPECT/CT-derived parameters were collected. Patients were stratified and randomly divided at a 7:3 ratio into a training cohort and a held-out internal test cohort. A two-phase modeling strategy was applied. Phase I models used clinical, laboratory, and conventional imaging features, whereas Phase II fusion models incorporated selected quantitative SPECT/CT parameters. Six algorithms, including LightGBM, XGBoost, CatBoost, Elastic Net, support vector machine (SVM), and Random Forest, were developed with five-fold cross-validation and grid search. Several inflammatory markers and quantitative SPECT/CT-derived parameters were higher in the PS group, whereas MRI-detected paravertebral abscess was more frequent in the STB group. In univariate ROC analysis, LBI SUVmax and LBI SUVmean showed the highest diagnostic performance among quantitative SPECT/CT parameters, with AUCs of 0.7681 and 0.7499, respectively. After incorporating these two parameters, the Phase II SVM fusion model achieved the best overall performance in the held-out internal test cohort, with an AUC of 0.8316 (95% CI 0.7097–0.9331), accuracy of 0.7885, sensitivity of 0.8333, specificity of 0.7273, F1 score of 0.8197, and Brier score of 0.1699. However, the Phase II Elastic Net model achieved the highest specificity of 0.7727 and precision of 0.8148. Compared with the Phase I SVM baseline model, the Phase II SVM fusion model showed an AUC increase of 0.0782, with a significant DeLong test result ( P = 0.028). Reclassification analysis showed NRI and IDI values of 0.18 and 0.09, respectively. Decision curve analysis indicated greater net benefit over most threshold probabilities between 0.16 and 0.59. SHAP analysis identified LBI SUVmax , LBI SUVmean , PCT, NLR, and SII as major contributors. Incorporating quantitative SPECT/CT-derived parameters into a multimodal machine learning framework improved the internally validated differentiation between STB and PS, with the Phase II SVM fusion model demonstrating the best overall performance and showing potential to support early, objective, and quantitative differential diagnosis.