Deep learning-based automatic classification of ischemic stroke subtype using diffusion-weighted images

Abstract

BACKGROUND: Accurate classification of ischemic stroke subtype is important for effective secondary prevention of stroke. We used diffusion-weighted imaging (DWI) and atrial fibrillation (AF) data to train a deep learning algorithm to classify stroke subtype. METHODS: Model training, validation, and internal testing were done in 2,988 patients with acute ischemic stroke from three stroke centers by using U-net for infarct segmentation and EfficientNetV2 for stroke subtype classification. Experienced vascular neurologists (n=5) determined stroke subtypes for external test datasets, while establishing a consensus for clinical trial datasets using the TOAST classification. Infarcts on DW images were automatically segmented using an artificial intelligence solution that we recently developed, and their masks were fed as inputs to a deep learning algorithm (DWI-only algorithm). Subsequently, another model was trained, with the presence or absence of AF included in the training as a categorical variable (DWI+AF algorithm). These models were tested: a) internally against the opinion of the labeling experts, b) against fresh external DWI data, and also c) against clinical trial DWI data acquired at a later date. RESULTS: In the training-and-validation datasets, the mean age was 68.0±12.5 (61.1% male). In internal testing, compared with the experts, the DWI-only algorithm and the DWI+AF algorithm respectively achieved moderate (65.3%) and near-strong (79.1%) agreement. In external testing, both algorithms again showed good agreements (59.3-60.7% and 73.7-74.0%, respectively). In the clinical trial dataset, compared with the expert consensus, percentage agreements and Cohen?s kappa were respectively 58.1% and 0.34 for the DWI-only algorithm vs. 72.9% and 0.57 for the DWI+AF algorithm. The corresponding values between experts were comparable (76.0% and 0.61) to the DWI+AF algorithm. CONCLUSIONS: Our deep learning algorithm trained on a large dataset of DWI (both with or without AF information) was able to classify ischemic stroke subtypes as accurately as a consensus of stroke experts.