Deep learning in acute ischemic stroke imaging: a systematic review of CT- and MRI-based segmentation, triage, and prognostic modeling.

backgroundAcute ischemic stroke is a time-critical neurological emergency in which imaging directly influences diagnosis, treatment eligibility, tissue-at-risk estimation, workflow prioritization, and outcome prediction. Deep learning has become increasingly prominent in stroke imaging; however, its clinical role differs substantially between computed tomography (CT) and magnetic resonance imaging (MRI).

objectiveThis PRISMA-guided systematic review evaluated peer-reviewed studies published between January 2023 and March 2026 that applied deep learning to acute ischemic stroke imaging.

methodsSearches across major biomedical, engineering, and clinical databases identified 91 eligible studies. Extracted data included imaging modality, dataset source, clinical task, model architecture, validation strategy, and reported performance metrics.

resultsMRI-based studies predominantly addressed lesion segmentation and tissue characterization using DWI, ADC, FLAIR, and multimodal MRI. In this setting, U-Net-derived architectures, self-configuring frameworks, attention-based models, and selected CNN-Transformer hybrids frequently achieved strong Dice and IoU values on benchmark datasets, particularly ISLES 2015 and ISLES 2022. CT-based studies followed a different clinical direction, with greater emphasis on emergency triage, hemorrhage exclusion, large-vessel occlusion detection, stroke classification, ASPECTS-related assessment, and CTP-based core-penumbra estimation. High retrospective accuracy or AUC values in CT studies should be interpreted cautiously, since image-level classification performance does not establish reliable voxel-level lesion delineation, especially on NCCT, where early ischemic changes remain subtle and difficult to segment.

conclusionAcross both modalities, technical progress remains ahead of clinical validation. External testing, patient-level separation, standardized reporting, scanner robustness, interpretability, and prospective workflow evaluation are still inconsistently addressed. The field now requires a shift from isolated benchmark performance toward clinically interpretable, externally validated, and prognosis-aware systems that can support real-time neuroradiology decision-making.