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

Area of Science:

• Neuroimaging analysis
• Artificial intelligence in medicine
• Stroke research

Background:

• Accurate segmentation of acute ischemic stroke (AIS) lesions is crucial for patient care.
• Manual segmentation is time-consuming and variable.
• Machine learning (ML) offers automated segmentation but lacks systematic performance evaluation.

Purpose of the Study:

• To systematically review and meta-analyze the performance of ML models for AIS lesion segmentation.
• To identify factors influencing model accuracy and robustness across different imaging modalities, architectures, and datasets.

Main Methods:

• Systematic review and meta-analysis of 101 studies (PRISMA 2020 guidelines).
• Searched PubMed, Scopus, Web of Science through March 2025 for ML-based AIS segmentation studies.
• Extracted data on study design, ML architecture, imaging, datasets, and performance metrics (Dice, AUC, accuracy, sensitivity, specificity).

Main Results:

• Deep learning models, especially U-Net variants (78%), showed high pooled Dice coefficient (0.84) and AUC (0.91).
• Multimodal MRI models outperformed single-modality CT.
• No significant association between lesion volume/sample size and Dice scores; performance trended higher with increasing clinical severity (mRS).

Conclusions:

• ML models, particularly deep learning on multimodal MRI, achieve high accuracy for AIS lesion segmentation.
• Model performance is robust across various datasets and lesion characteristics.
• Need for methodological standardization and external validation for clinical integration.