Predicting the prognosis of symptomatic intracranial atherosclerotic stenosis (sICAS) patients using deep learning...

Symptomatic intracranial atherosclerotic stenosis (sICAS) is a leading cause of stroke recurrence. This study aimed to develop a deep learning model based on high-resolution vessel wall imaging (HR-VWI) to improve recurrence prediction, identify high-risk patients, and guide clinical intervention.

We retrospectively collected HR-VWI data from 363 patients with sICAS across two medical centres. Centre 1 (n = 254) served as the training cohort, and centre 2 (n = 109) served as the external validation cohort. Three deep learning models-ResNet50, DenseNet169, and Vision Transformer (ViT)-were used to extract features from both 2D and 3D images of culprit plaques. In addition, radiomics-based machine learning models were constructed using manually extracted features. The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was applied for feature selection, and a Naive Bayes classifier was used to predict the risk of stroke recurrence. Model performance was evaluated using the area under the curve (AUC), calibration curves, and decision curve analysis (DCA).

The 3D-ResNet50 (AUC = 0.780, 95% confidence interval [CI]: 0.638-0.92) and 3D-DenseNet169 (AUC = 0.780, 95% CI: 0.641-0.919) models significantly outperformed 2D models: 2D-ResNet50 (AUC = 0.574, 95% CI: 0.416-0.719), 2D-DenseNet169 (AUC = 0.660, 95% CI: 0.533-0.788), and radiomics (AUC = 0.698, 95% CI: 0.579-0.810). Delong's test confirmed the significance of these differences. Calibration and DCA curves further underscored the 3D models' clinical value.

The 3D deep learning model based on HR-VWI offers superior prediction of sICAS recurrence risk compared to 2D models and radiomics, aiding clinical decision-making and high-risk patient identification.