Machine learning models of segmentation in acute ischemic stroke: a systematic review and meta-analysis

Accurate segmentation of acute ischemic stroke (AIS) lesions on neuroimaging is essential for diagnosis, treatment decision-making, and prognostication. Manual methods are limited by time and variability. Machine learning (ML), especially deep learning, has emerged as a powerful tool for automated lesion segmentation, yet a systematic synthesis of model performance, methodological rigor, and clinical applicability remains lacking.

To systematically review and quantitatively evaluate the performance of ML-based segmentation models for AIS using a meta-analytic approach, and to identify factors associated with model accuracy and robustness across imaging modalities, architectures, and datasets.

We conducted a systematic review and meta-analysis in accordance with PRISMA 2020 guidelines. Comprehensive searches were performed in PubMed, Scopus, and Web of Science databases through March 2025. Eligible studies included those reporting on machine learning (ML)-based segmentation of acute ischemic stroke (AIS) lesions on CT or MRI and providing quantitative performance metrics (e.g., Dice, sensitivity, specificity, AUC). Data were systematically extracted on study design, ML architecture, imaging modality, dataset size and composition, and segmentation performance. Random-effects meta-analyses were conducted using inverse-variance weighting, with logit transformation applied to bounded metrics to stabilize variance. Between-study heterogeneity was assessed using the I² statistic and Cochran's Q-test. Meta-regression analyses explored the influence of covariates such as lesion volume, sample size, and stroke severity (mRS), while subgroup analyses examined performance variations by imaging modality and model type. Visualizations included forest plots, funnel plots, bubble plots, and correlation matrices, generated in Python using standardized meta-analysis and statistical libraries.

Out of 4755 screened records, 101 studies met the inclusion criteria. Deep learning approaches, especially U-Net variants, dominated the field (78%). The pooled Dice coefficient was 0.84 (I² = 0%), indicating high and consistent segmentation accuracy. Additional pooled estimates were AUC 0.91, accuracy 0.89, sensitivity 0.85, and specificity 0.93. Models using multimodal MRI inputs outperformed single-modality CT models. Meta-regressions revealed no significant association between lesion volume or sample size and Dice scores, though segmentation performance trended higher with increasing clinical severity (mRS). Specificity showed weak correlation with recall and Dice, indicating potential trade-offs in model optimization.

ML-based segmentation models for AIS demonstrate high accuracy, especially when using multimodal MRI and deep learning architectures. Performance is robust across varying dataset sizes and lesion characteristics. However, heterogeneity in study designs and reporting standards underscores the need for methodological harmonization and external validation. These findings inform future model development and integration into clinical AIS workflows.