Development and validation of a DWI-based intra- and peri-infarction radiomics model for predicting neurological deterioration in acute ischemic stroke

Abstract

Introduction Early prediction of neurological deterioration (ND) in acute ischemic stroke (AIS) remains challenging. We aimed to develop a diffusion-weighted imaging (DWI)-based radiomic model that integrates intra- and peri-infarction radiomic signatures with clinical characteristics to improve ND prediction. Methods This retrospective study included 767 patients with anterior circulation AIS (training, n = 537; test, n = 230) who underwent DWI within 3 days of onset. Infarct cores were manually segmented, and peri-infarction regions (1–3 mm annular expansions) were generated. Radiomics features were extracted from all regions. Significant clinical predictors and radiomics features were selected to construct four models: clinical, intra-infarction radiomics, peri-infarction radiomics (1–3 mm), and a combined model. Performance was assessed using area under the curve (AUC), sensitivity, specificity, calibration, and decision curve analysis (DCA). Results In total, 767 patients were included in the analysis. In multivariate analysis, glucose level was an independent clinical predictor of ND [odds ratio (OR) = 1.003; 95% confidence interval (95% CI), 1.001–1.006; p < 0.05]. Overall, 1,561 radiomic features were extracted from the intra-infarction region. Ten radiomics features were deemed valuable for dimensionality reduction and selection. Similarly, 1,561 radiomic features were extracted from the Peri2mm region, and 16 radiomics features were included. Among the intra-infarction radiomics models, the XGBoost model showed the best predictive efficiency, with AUCs of 0.911 and 0.898 for the training and test cohorts, respectively. The 2 mm peri-infarction model performed best among all peri-infarction models (AUCs: 0.854 for training, 0.850 for test). The combined model (intra-infarction + 2 mm peri-infarction + clinical feature) achieved the highest AUC values of 0.922 (training) and 0.911 (test), accompanied by favorable calibration and stable clinical net benefit. Although the combined model yielded a perfect sensitivity of 1.000 in the test cohort (95% Wilson CI: 0.867–1.000), this finding should be interpreted cautiously due to the small number of ND events (n = 25). Further pairwise comparison revealed that the combined model possessed significantly better discriminative ability than the clinical model and the 1 mm peri-infarction model in the test cohort (p < 0.05). However, the improvement over intra-infarction radiomics alone did not reach statistical significance in the test cohort (DeLong p = 0.179), and no significant superiority was observed versus the Peri2mm and Peri3mm models either. Discussion Integrating DWI-derived intra-infarction radiomics, 2 mm peri-infarction signatures and clinical indicator achieves optimal AUC discrimination within the training cohort, though its statistical advantage over isolated intra-infarction radiomics could not be validated in the independent test set. This multimodal nomogram provides a non-invasive imaging tool to assist ND risk stratification for acute ischemic stroke patients, while larger multi-center cohorts are required to confirm its universal incremental predictive value and stable detection performance.