Related Experiment Video
Updated: Sep 19, 2026
Time-Resolved, Dynamic Computed Tomography Angiography for Characterization of Aortic Endoleaks and Treatment Guidance via 2D-3D Fusion-Imaging
Published on: December 9, 2021
Radiomics and deep learning for intracranial aneurysm rupture-status assessment on CTA/MRA: a multicenter
1Department of Neurosurgery, The Affiliated Hospital of Southwest Medical University, Luzhou, Sichuan, China.
Background:
Rupture of intracranial aneurysms leads to aneurysmal subarachnoid hemorrhage with severe adverse outcomes. Current retrospective imaging studies merely differentiate existing ruptured and unruptured aneurysms instead of forecasting prospective rupture risk. To address this limitation, we constructed a multimodal fusion model combining handcrafted radiomics and deep learning embeddings from routine CTA/MRA for aneurysm rupture discrimination and performed internal validation.
Methods:
This multicenter retrospective study screened 520 patients with 572 saccular aneurysms. After eligibility and imaging quality control, 400 patients with 400 index aneurysms were enrolled. Patients were divided into training, validation and test sets (70/15/15) with balanced ruptured/unruptured aneurysm ratios. A total of 1,316 radiomic features were extracted from aneurysm segmentations; stable, non-redundant predictors (n=42) were retained via filtering and LASSO regression. A 3D ResNet produced 512-dimensional deep embeddings, compressed to 128 components by PCA. We fused radiomic and deep features as a 170-dimensional input for an XGBoost classifier. Model performance was assessed via discrimination, calibration and decision-curve analysis.
Results:
The fusion model reached an AUC of 0.928, accuracy 0.889, sensitivity 0.902, specificity 0.872, PPV 0.886 and NPV 0.889 on the test set, superior to single radiomic and deep learning models. The model exhibited excellent calibration (slope=0.98, Brier score=0.082). Decision-curve analysis demonstrated greater net clinical benefit across probability thresholds of 0.10-0.70.
Discussion:
Integrating reproducible radiomic metrics and volumetric deep embeddings significantly improves the discrimination of aneurysm rupture status. Nevertheless, all rupture labels were determined at initial presentation, and external independent validation was absent. Thus, this pipeline is only an internally validated methodological tool for imaging phenotype analysis, rather than a ready-to-use clinical model for predicting future aneurysm rupture.
