Integrating Clinical Data and Radiomics and Deep Learning Features for End-to-End Delayed Cerebral Ischemia Prediction on Noncontrast CT

BACKGROUND AND PURPOSE: Delayed cerebral ischemia is hard to diagnose early due to gradual, symptomless development. This study aimed to develop an automated model for predicting delayed cerebral ischemia following aneurysmal SAH on NCCT.

MATERIALS AND METHODS: This retrospective study included 400 patients with aneurysmal SAH (156 with delayed cerebral ischemia) who underwent NCCT. The study used ATT-Deeplabv3+ for automatically segmenting hemorrhagic regions using semisupervised learning. Principal component analysis was used for reducing the dimensionality of deep learning features extracted from the average pooling layer of ATT-DeepLabv3+. The classification model integrated clinical data, radiomics, and deep learning features to predict delayed cerebral ischemia. Feature selection involved Pearson correlation coefficients, least absolute shrinkage, and selection operator regression. We developed models based on clinical features, clinical-radiomics, and a combination of clinical, radiomics, and deep learning. The study selected logistic regression, Naive Bayes, Adaptive Boosting (AdaBoost), and multilayer perceptron as classifiers. The performance of segmentation and classification models was evaluated on their testing sets using the Dice similarity coefficient for segmentation, and the area under the receiver operating characteristic curve (AUC) and calibration curves for classification.

RESULTS: The segmentation process achieved a Dice similarity coefficient of 0.91 and the average time of 0.037 s/image. Seventeen features were selected to calculate the radiomics score. The clinical-radiomics–deep learning model with multilayer perceptron achieved the highest AUC of 0.84 (95% CI, 0.72–0.97), which outperformed the clinical-radiomics model (P = .002) and the clinical features model (P = .001) with multilayer perceptron. The performance of clinical-radiomics–deep learning model using AdaBoost was significantly superior to its clinical-radiomics model (P = .027). The performance of the clinical-radiomics–deep learning model and the clinical-radiomics model with logistic regression notably exceeded that of the model based solely on clinical features (P = .028; P = .046). The AUC of the clinical-radiomics–deep learning model with multilayer perceptron (P < .001) and the clinical-radiomics model with logistic regression (P = .046) were significantly higher than the clinical model with logistic regression. Of all models, the clinical-radiomics–deep learning model with multilayer perceptron showed best calibration.

CONCLUSIONS: The proposed 2-stage end-to-end model not only achieves rapid and accurate segmentation but also demonstrates superior diagnostic performance with high AUC values and good calibration in the clinical-radiomics–deep learning model, suggesting its potential to enhance delayed cerebral ischemia detection and treatment strategies.