Related Experiment Video
Updated: Aug 6, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Artificial Intelligence in Ischemic Stroke Lesion Segmentation: A Narrative Review of Deep Learning Methods, Clinical
Taofeeq Oluwatosin Togunwa1,2, Theophilus Olayiwola3, Halleluyah Darasinmi Oludele3
1Department of Learning Health Sciences, University of Michigan Medical School, Ann Arbor, MI, USA.
Journal of Imaging Informatics in Medicine
|July 17, 2026
Summary
Deep learning (DL) for ischemic stroke lesion segmentation shows promise, particularly for MRI, with improving accuracy. CT segmentation faces challenges but is advancing, highlighting the need for further validation and deployment.
Area of Science:
- Medical Imaging and Artificial Intelligence
- Neuroscience and Neurology
- Computer Science and Machine Learning
Background:
- Ischemic stroke lesion segmentation is critical for timely management, treatment selection, and prognostication.
- Deep learning (DL) offers potential to automate and standardize lesion delineation in neuroimaging.
- Recent advancements in DL architectures are being explored to improve stroke lesion segmentation accuracy.
Purpose of the Study:
- To conduct a narrative review of deep learning-based ischemic stroke lesion segmentation studies.
- To synthesize findings on imaging modalities, DL architectures, and performance metrics (Dice Similarity Coefficient [DSC]).
- To identify trends and challenges in DL stroke lesion segmentation from 2020 to 2025.
Main Methods:
- A narrative review of studies published between 2020 and 2025 was performed.
- Literature search conducted across PubMed, Google Scholar, Scopus, and IEEE Xplore.
- Extracted data included imaging modality, DL architecture, segmentation targets, and reported DSC performance.
Main Results:
- U-Net architectures and variants dominated, with residual and attention mechanisms enhancing performance.
- MRI segmentation (especially DWI/ADC) showed high performance (DSC > 0.80), with advanced models approaching 0.90.
- CT/NCCT segmentation performance was more variable (DSC ~0.35-0.65), but newer hybrid and ensemble methods neared 0.80.
Conclusions:
- Deep learning for stroke lesion segmentation is maturing, showing strong potential for clinical use, especially with MRI.
- CT segmentation faces challenges due to subtle early changes and generalization issues, requiring further development.
- Prospective, multicenter validation and scalable deployment are crucial for real-world impact and equitable access.
