Background: Pediatric arterial ischemic stroke (PAIS) is a significant cause of acquired brain injury and mortality in childhood. Prediction of stroke risk and outcome is complex and depends on multiple heterogenous factors including lesional and non-lesional neuroimaging features. However, manual lesion segmentation of acute PAIS is labor-intensive and can produce inconsistent results, highlighting the need for scalable, automated methods. Research in automated segmentation models focuses primarily on adults, and these models often fail to generalize to pediatric cases. The purpose of this study was to develop a robust, automated segmentation model for acute PAIS using a deep learning approach tailored to this patient population. Methods: The dataset included clinically acquired diffusion weighted imaging (DWI), b1000, and apparent diffusion coefficient (ADC) MRI images in PAIS (N=142) participants in an institutional registry, along with radiologist-annotated ground truth masks. The network used was an nnU-Netv2 model, which self-configures its architecture to the dataset's specific characteristics, removing the need for manual configuration. The model was trained using 5-fold cross-validation, where the dataset was split into five subsets and each fold was used once as a validation set while the remaining folds were used for training. For final evaluation, predictions from all five folds were combined using ensemble inference. Performance was evaluated using Dice score, Intersection over Union (IoU), and qualitative inspection of the segmentations. Results: The model achieved a mean Dice score of 0.71 and an IoU of 0.55 across test cases, with consistent lesion boundary delineation. A radiologist-supported qualitative review confirmed alignment with vascular territories and high agreement with ground truth annotations. Conclusions: This study demonstrates the successful development of a highly accurate and robust automated segmentation model for PAIS. Utilizing a self-configuring nnU-Netv2 framework, the model performs strongly on a rare patient population and aligns with radiologist annotations. While human-in-the-loop review remains important for refining edge cases, these automated predictions significantly accelerate the segmentation process and reduce the workload for radiologists. This work represents a crucial step toward scalable, automated tools that can significantly reduce the workload and improve the prediction of risk and outcome in PAIS.
Alex et al. (2026) studied this question.