Multimodal deep learning model based on CT perfusion-defined infarct core for predicting parenchymal hematoma type 2...

To develop and validate a multimodal deep learning model based on a CT perfusion (CTP)-defined infarct core for predicting parenchymal hematoma type 2 (PH2) after endovascular therapy (EVT) in patients with acute ischemic stroke (AIS).

In this dual-center retrospective study, 487 patients with anterior-circulation large-vessel-occlusion AIS who underwent EVT were included, with 219 assigned to the training cohort and 268 to an independent external validation cohort. The infarct core was defined using CTP and transferred to non-contrast CT (NCCT) to generate three-dimensional image patches. A score-based baseline model (Model.Score) was constructed using logistic regression with ASPECTS as the predictor. A pure imaging model based on a 3D DenseNet architecture (Model.Core) was developed using NCCT images and infarct-core masks as dual-channel inputs. In addition, a conventional clinical-imaging model (Model.Clinical-Imaging) was constructed using logistic regression based on ASPECTS and baseline clinical variables. A multimodal fusion model (Model.Fusion) was subsequently developed by integrating infarct-core imaging features with baseline clinical variables and dual-phase CT angiography (CTA)-derived scores. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration analysis, decision curve analysis, and statistical comparison with DeLong, net reclassification improvement, and integrated discrimination improvement tests. Model interpretability was evaluated using gradient-weighted class activation mapping and permutation importance analysis.

Model.Fusion achieved the best predictive performance, with AUCs of 0.906 in the training cohort and 0.845 in the external validation cohort, exceeding those of the score-based model, the conventional clinical-imaging model, and Model.Core. The fusion model also showed better calibration and greater clinical net benefit. Interpretability analyses demonstrated predominant model attention to low-density regions within the infarct core, while dual-phase CTA-derived variables and baseline NIHSS score were identified as major predictors.

CTP-defined infarct core-based deep learning enables effective prediction of PH2 after EVT in AIS. Integration of imaging, clinical, and dual-phase CTA features further improves predictive performance and may facilitate individualized risk stratification before treatment.