A Deep-Learning Model for Stroke Subtype Identification in Moyamoya Disease Based on Digital Subtraction Angiography.

Abstract

Moyamoya disease (MMD) may present with ischemic or hemorrhagic stroke, with differing prognosis and management. Early identification of the likely stroke subtype could support individualized therapeutic planning. We developed an interpretable digital subtraction angiography (DSA)–based deep-learning model to classify stroke subtype in adult MMD. This retrospective study included 163 adults with MMD (98 ischemic, 65 hemorrhagic) who underwent DSA between January and December 2016. Images from predefined vessels and projection planes were preprocessed and augmented. Convolutional neural network (CNN) classifiers were trained using single-vessel images and combined-vessel images. Data were split into a training set (80%, n = 130) and an independent test set (20%, n = 33). Model interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM++). Baseline characteristics were comparable between the training and test sets. Among single-vessel models, sagittal vertebral artery (VA) images achieved the best performance (F1 = 0.929; accuracy = 0.939), comparable to sagittal external carotid artery (ECA) images. For the same vessel, sagittal projections consistently outperformed coronal views. A combined model incorporating sagittal internal carotid artery (ICA) and sagittal VA images also achieved optimal performance (F1 = 0.929; accuracy = 0.939). Grad-CAM + + highlighted discriminative regions along the ICA and VA, consistent with clinical assessment. A DSA-based CNN can accurately distinguish ischemic versus hemorrhagic presentation in adult MMD. Posterior circulation features—particularly the VA—and sagittal projections provide strong predictive signal, supporting the translational relevance of posterior circulation morphology for hemorrhagic MMD.