Machine learning model based on CT radiomics for the differential diagnosis of brucellar and tuberculous spondylitis: a model construction and validation study

Le résumé fourni par la source

Tuberculous spondylitis (TS) and brucellar spondylitis (BS) present with overlapping clinical and conventional imaging features, rendering their differential diagnosis challenging. While prior work demonstrated the feasibility of machine learning for TS versus BS differentiation using MRI-based features, the added value of CT-based radiomics—which provides complementary bone microstructural information—integrated with clinical characteristics has not been systematically investigated. This study aimed to develop and internally validate a CT radiomics-based machine learning model combining radiomic features with clinical variables for differentiating TS from BS in surgically confirmed patients. This single-center retrospective study enrolled 327 surgically treated and pathologically confirmed patients (259 TS, 68 BS) who underwent CT examination (January 2017 – December 2022). Radiomic features ( n = 1,051) were extracted from sagittal CT images using PyRadiomics in compliance with IBSI standards. After ComBat harmonization for scanner effects, feature selection employed minimum redundancy maximum relevance (mRMR) ranking followed by LASSO regression (λ = λ_1se). Five classifiers—logistic regression (LR, prespecified as primary), support vector machine, random forest, XGBoost, and k-nearest neighbors—were evaluated for clinical, radiomics, and combined models under stratified 10-fold nested cross-validation. All preprocessing, feature selection, and hyperparameter tuning were confined to training folds. Model performance was assessed using area under the ROC curve (AUC), precision–recall AUC, sensitivity, specificity, calibration, and decision-curve analysis. The LR-based combined model achieved a mean cross-validated AUC of 0.882 (95% CI, 0.837–0.927), sensitivity of 0.867, and specificity of 0.671. The radiomics-alone model yielded an AUC of 0.847 (95% CI, 0.797–0.897), and the clinical model yielded an AUC of 0.786 (95% CI, 0.726–0.846). The combined model provided a modest but consistent improvement over the radiomics model (ΔAUC = 0.035; DeLong P = 0.068, adjusted P = 0.136). Among five classifiers, LR demonstrated the smallest training-validation AUC decrement (ΔAUC = 0.046), indicating favorable stability relative to ensemble methods. The combined model’s specificity of 0.671 corresponds to a false-positive rate of 0.329, indicating that approximately one-third of TS cases may be misclassified as BS, a limitation addressed in the Discussion. The combined CT radiomics-clinical model based on logistic regression demonstrates moderate-to-good discriminative performance for differentiating TS from BS in surgically confirmed patients. However, its imperfect specificity, the single-center retrospective design, and the restriction to surgically treated patients limit immediate generalizability. Prospective multi-center external validation and integration of brucellosis-specific epidemiological exposure history are needed before clinical translation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.