Artificial Intelligence in Ischemic Stroke Lesion Segmentation: A Narrative Review of Deep Learning Methods, Clinical Utility, and Future Directions.
Authors
Affiliations (12)
Affiliations (12)
- Department of Learning Health Sciences, University of Michigan Medical School, Ann Arbor, MI, USA.
- Michigan Institute for Data and AI in Society, University of Michigan, Ann Arbor, MI, USA.
- College of Medicine, University of Ibadan, Ibadan, Nigeria.
- Department of Radiology, University College Hospital, Ibadan, Nigeria.
- Department of Neurosurgery, Duke University Health System, Durham, NC, USA.
- Duke Global Health Institute, Durham, NC, USA.
- Duke Global Neurosurgery and Neurology, Durham, NC, USA.
- Department of Neurology, Duke University Health System, Durham, NC, USA.
- Department of Radiology, University of Ibadan, Ibadan, Nigeria.
- Department of Neurosurgery, Duke University Health System, Durham, NC, USA. [email protected].
- Duke Global Health Institute, Durham, NC, USA. [email protected].
- Duke Global Neurosurgery and Neurology, Durham, NC, USA. [email protected].
Abstract
Ischemic stroke management is time-sensitive, and lesion segmentation supports treatment selection, prognostication, and reproducible quantification. Deep learning (DL) aims to accelerate and standardize lesion delineation to augment neuroimaging workflows. We conducted a narrative review of DL-based ischemic stroke lesion segmentation studies published from 2020 to 2025. PubMed, Google Scholar, Scopus, and IEEE Xplore were searched; ~ 500 records were identified, and 40 full-text studies were included after screening. We extracted imaging modality, architecture, segmentation targets, and reported performance (primarily Dice similarity coefficient [DSC]) for descriptive synthesis; exploratory LOESS curves were used only to visualize broader temporal trends and were not interpreted inferentially. U-Net backbones and variants remained dominant, with gains from residual and attention mechanisms and standardized pipelines. On MRI (especially DWI/ADC), many studies reported DSC > 0.80; residual/attention U-Nets reached ~ 0.87 and nnU-Net achieved ~ 0.80-0.82. Since 2023, transformer-based and ensemble approaches have increasingly appeared among top-performing MRI models, with multisite DWI reports approaching ~ 0.90. CT/NCCT segmentation was more variable, with typical research DSC ~ 0.35-0.65, but late-period hybrid CNN-transformer, ensemble, and multimodal approaches reported DSC approaching ~ 0.8, and NCCT tools report higher performance. Overall trends suggested gradual improvement for MRI and a U-shaped trajectory for CT. DL stroke lesion segmentation is maturing toward clinical viability, most convincingly for MRI, while CT applications remain constrained by subtle early ischemic change and generalization challenges. Many studies remain retrospective and may not reflect real-world performance and access constraints, reinforcing the need for prospective, multicenter validation and scalable deployment pathways to enable equitable impact.