Development and internal validation of CTA-based radiomics models for differentiating symptomatic from asymptomatic unruptured spontaneous intracranial vertebrobasilar artery dissection.
Authors
Affiliations (1)
Affiliations (1)
- Department of Neurosurgery, Beijing Neurosurgical Institute and Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
Abstract
Spontaneous intracranial vertebrobasilar artery dissection (sIVBD) has heterogeneous clinical presentations, and conventional clinical and morphologic assessment may not fully characterize lesion heterogeneity. We aimed to develop and internally validate CTA-based radiomics models for differentiating symptomatic from asymptomatic unruptured sIVBD and to evaluate the incremental contribution of radiomics beyond conventional clinical and CTA features. This retrospective single-center study included 173 patients with unruptured sIVBD, comprising 56 symptomatic and 117 asymptomatic patients. Patients were stratified into a development cohort (<i>n</i> = 138) and an independent internal validation cohort (<i>n</i> = 35). Radiomic features were extracted from subtraction CTA images after three-dimensional lesion segmentation. Feature reproducibility was assessed using inter-reader intraclass correlation coefficients. Four models were developed: a conventional CTA model (M1), a clinical plus conventional CTA model (M2), a radiomics XGBoost model (M3), and a combined model integrating M2 predictors with a cross-fitted radiomics probability (M4). Primary development-cohort evaluation used five-times-repeated five-fold stratified nested cross-validation. Discrimination, calibration, Brier score, and decision-curve analysis were assessed. Of 851 radiomic features, 722 (84.8%) met the prespecified reproducibility criterion. In patient-averaged out-of-fold evaluation, AUCs were 0.630 for M1, 0.681 for M2, 0.676 for M3, and 0.723 for M4. Compared with M2, M4 increased AUC by 0.0416 (95% CI, -0.0130 to 0.0962; <i>p</i> = 0.136) and reduced the Brier score by 0.0101 (95% bootstrap CI, -0.0208 to 0.0006). Across 25 outer folds, M4 had the highest mean AUC (0.715) and the lowest between-fold variability. In the independent internal validation cohort, AUCs were 0.814, 0.807, 0.803, and 0.966 for M1 through M4, respectively. A combined model integrating clinical variables, conventional CTA features, and cross-fitted radiomics information showed the best overall performance for differentiating symptomatic from asymptomatic unruptured sIVBD. Radiomics may provide complementary quantitative information beyond conventional assessment.