Automated imaging analysis in the era of stroke: a narrative review

Introduction

Rationale/background

Automated imaging analysis (AIA) refers to advanced technology, specifically the use of artificial intelligence (AI) software to analyze medical images. These images may come from computed tomography (CT) or magnetic resonance imaging (MRI) scans, along with their associated imaging series. Its application in medicine, especially in Radiology, is revolutionizing clinical practice, with changes occurring in everyday practice. Machine learning (ML) and deep learning (DL) are subfields of AI. ML works by training algorithms on large datasets to identify patterns, make predictions and improvement as they process more data over time (1). DL works by using artificial neural networks (ANNs) to automatically learn patterns and features from large datasets (1). In Neuroradiology, where there is often a need for immediate diagnosis and management, AIA plays a crucial role in assisting with the diagnosis. Imaging interpretation can be time-consuming and subject to variability among radiologists, potentially delaying treatment decisions. Advanced AI-driven algorithms aim to improve sensitivity, and specificity in identifying key findings, reducing missed diagnoses and ensuring better public health (2).

Stroke is the third leading cause of deaths (after ischemic heart disease and coronavirus disease 2019) with 11.9 million new cases occurring annually (3). It is classified into ischemic stroke, which accounts for about 87% of cases, and hemorrhagic stroke, which represents the remaining 13% (4). Ischemic strokes are further classified based on the size and location of the affected vessel. Large vessel occlusions (LVOs) involve major intracranial arteries and are typically treated with both intravenous thrombolysis (IVT) and mechanical thrombectomy (MT), the latter being the gold standard for eligible patients (5,6). In contrast, in medium vessel occlusions (MeVOs) and distal vessel occlusions (DVOs) treatment options remain less well-defined. IVT remains the primary therapy, but MT’s role in MeVOs and DVOs is still under investigation (7,8). Given the significant burden of stroke-related disability, optimizing therapeutic strategies remains a crucial area of research.

CT and MRI have long been standard imaging tools for stroke evaluation, each offering specific advantages and limitations. Non-contrast CT (NCCT) is widely available, rapid, and essential for excluding intracranial hemorrhage in the acute setting; however, it has limited sensitivity for early ischemic changes and may miss small or subtle infarcts in the hyperacute phase. CT angiography (CTA) enables the visualization of intracranial and extracranial vessel occlusions and stenoses but requires iodinated contrast agents and carries a risk of nephrotoxicity, particularly in patients with impaired renal function. CT perfusion (CTP) allows for the estimation of infarct core and penumbra by quantifying parameters such as cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT). However, it is sensitive to technical variability, such as incorrect arterial input function placement, and may be affected by patient motion, contrast timing, and radiation exposure. On the other hand, MRI offers superior sensitivity in detecting early ischemic lesions, particularly through diffusion-weighted imaging (DWI), which can identify acute infarcts within minutes of onset. Nonetheless, MRI is less accessible in emergency settings, has longer acquisition times, and is contraindicated in patients with certain implants or severe claustrophobia. MR angiography (MRA) provides non-invasive vessel assessment like CTA but may have lower spatial resolution and can sometimes overestimate the degree of stenosis, especially in time-of-flight sequences (9).

Objectives

While numerous reviews have explored AI applications in stroke care, most focus narrowly on early diagnosis or LVOs, often neglecting the broader clinical context and the continuum of stroke management. Critically, there is limited literature specifically addressing the role of AIA in medium and distal vessel occlusions (MeVOs/DMVOs)—areas gaining increasing clinical attention but still underrepresented in AI research. This review aims to fill that gap by offering an integrated overview of AIA applications across all stages of acute ischemic stroke (AIS) care, including less-discussed domains like MeVOs/DMVOs, hemorrhagic transformation, outcome prediction, and rehabilitation. Unlike existing reviews, it also examines the clinical utility, ethical challenges, and real-world limitations of AIA tools, providing a more practical and balanced perspective on their current and future role in stroke management. We present this article in accordance with the Narrative Review reporting checklist (available at https://jni.amegroups.com/article/view/10.21037/jni-25-15/rc).

Methods

A literature search was conducted to identify relevant studies published between January 2019 and January 2025 using the databases PubMed, Google Scholar, and Scopus (Table 1). The search strategy employed a combination of keywords and Medical Subject Headings (MeSH) terms including artificial intelligence, machine learning, deep learning, stroke imaging, ischemic stroke, hemorrhagic stroke, large vessel occlusion, medium vessel occlusion, distal vessel occlusion, perfusion analysis, brain segmentation, stroke diagnosis, stroke treatment, stroke outcome, and post-stroke rehabilitation. Non-English studies were excluded. The aim was to capture the breadth of AI applications across the stroke care continuum, including diagnostic enhancement, therapeutic decision support, prognostic modeling, and post-stroke recovery monitoring. The search yielded a diverse range of literature, including original articles, reviews, and meta-analyses. Priority was given to recent publications. Key findings were extracted and synthesized to provide a structured overview of current trends and advancements in AI-based imaging analysis in stroke care.

Discussion

Role and applications of AIA in stroke

AIA has multiple roles in stroke care, with applications beginning from the moment a stroke is suspected and extending through every stage, including diagnosis, treatment planning, outcome evaluation and post-stroke rehabilitation.

Role and applications of AIA in stroke diagnosis

To begin with, AIA plays a pivotal role in the early detection of stroke by significantly enhancing both the speed and accuracy of identifying critical radiological findings. This is particularly valuable in acute stroke care, where rapid decision-making can substantially impact patient outcomes. By minimizing interpretation time and reducing human error, AIA supports timely initiation of appropriate therapeutic interventions, especially in time-sensitive scenarios such as hyper AIS or intracranial hemorrhage (10).

In hemorrhagic stroke cases, AIA enables prompt detection and localization of intracranial bleeding, facilitating the crucial distinction between hemorrhagic and ischemic etiologies, an essential step in guiding appropriate treatment strategies. In ischemic stroke, AIA algorithms can identify early parenchymal changes, delineate the ischemic core, and estimate the surrounding penumbral tissue. These capabilities directly support the assessment of the Alberta Stroke Program Early CT Score (ASPECTS), a widely used grading system for evaluating early ischemic changes on NCCT, as illustrated in Figure 1.

Figure 1 NCCT scan showing an automated ASPECTS assessment. A hypointense lesion is visible in the left frontoparietal lobe and the insular cortex. The ischemic areas are highlighted within a red frame, automatically mapped according to the ASPECTS criteria. ASPECTS, Alberta Stroke Program Early CT Score; NCCT, non-contrast computed tomography.

Adamou et al. conducted a meta-analysis evaluating the diagnostic performance of automated ASPECTS algorithms using data from 11 studies encompassing 1,976 patients. Their findings demonstrated good concordance between automated and expert ASPECTS ratings, with an intraclass correlation coefficient (ICC) of 0.72. Furthermore, in the detection of early ischemic changes, the automated tools performed comparably—or in some cases superiorly—to expert radiologists, with a moderate correlation reported between individual physician and automated assessments (ICC: 0.54) (11). These results underscore the potential of AIA to augment clinical decision-making by offering reliable and consistent assessments, particularly in settings where expert neuroradiological input may be limited or delayed.

AIA is important for cerebral lesion segmentation, helping to accurately identify and measure different cerebral parts, especially in difficult cases with early stroke signs which are difficult to distinguish. Using advanced AI software, such as DL models, the brain can be divided into specific anatomical areas. Qiu et al. developed an ML method to automatically detect and measure infarcted brain tissue in AIS patients using NCCT scans. Trained on 157 patients and tested on 100, it showed strong agreement with MRI results, with an average stroke size difference of 11 mL. This method can aid in detecting ischemic lesions on CT scans, improving diagnosis and treatment planning (12). Öman et al., explored the use of a 3D convolutional neural network (CNN) to detect AIS from CTA source images. CTA from 60 patients were analyzed, with 30 used for training and 30 for testing, based on manually segmented lesions. The model successfully identified stroke lesions with 0.93 sensitivity and 0.82 specificity, achieving a 0.93 AUC and a 0.61 Dice coefficient. Performance improved when comparing both brain hemispheres, while adding NCCT had little effect. The findings suggest that CNN-based software can accurately detect ischemic stroke from CTAs, with hemispheric comparison helping reduce false positives (13). Gheibi et al., introduced a CNN-Res, a DL model designed for automatic segmentation of ischemic stroke lesions from multimodal MRI scans, addressing the limitations of manual segmentation, which is time-consuming and dependent on neurologist expertise. The model follows a U-shaped architecture with residual units to improve training and extract complex image details, while a bottleneck strategy reduces parameters and training time. Evaluated on two datasets, CNN-Res achieved 85.43% Dice coefficient on data from Tabriz University and 79.23% on the SPES 2015 dataset, demonstrating competitive accuracy. Overall, CNN-Res offers an efficient and accurate method for stroke lesion segmentation, aiding neuroradiologists in diagnosis (14). In general, these advancements underscore the significant potential of AI-driven cerebral lesion segmentation to improve the accuracy, reproducibility, and efficiency of stroke imaging analysis, ultimately contributing to enhanced clinical decision-making and patient care.

Perfusion imaging remains a cornerstone in the evaluation of AIS, providing critical information for the infarct core and penumbra (Figure 2). One of the most widely adopted approaches involves the use of AI to automate the selection of the arterial input function (AIF) and venous output function (VOF), thereby reducing operator dependence and minimizing partial volume effects. Automated analysis typically quantifies CBF, with thresholds such as CBF <30% of normal values commonly used to estimate the ischemic core. In parallel, time-to-maximum (T-max) thresholds—most often T-max >6 seconds—are applied to delineate hypoperfused tissue at risk of infarction. The accuracy of these estimates can be compromised by improper placement of AIF or VOF, especially when the AIF is selected distal to an occlusion. Additionally, a range of technical and physiological factors—including imaging parameters, contrast injection characteristics, and patient-specific vascular conditions—can influence perfusion metrics. Therefore, careful interpretation of perfusion maps and time-attenuation curves is essential to ensure accurate tissue characterization and appropriate patient selection for reperfusion therapies (15,16).

Figure 2 Automated MR perfusion analysis. In the left parietal lobe, the ischemic core is highlighted in pink, while the penumbra is shown in green. MR, magnetic resonance; rCBF, relative cerebral blood flow; rCBV, relative cerebral blood volume; T-max, time-to-maximum; TTP, time-to-peak.

Moreover, numerous AI-based software tools have been developed to estimate the ischemic core. Kasasbeh et al. used ANN to predict the ischemic core in acute stroke patients based on CTP data, comparing results to DWI as the gold standard. The ANN using only CTP data predicted the ischemic core with an average error of 13.8 mL and achieved an area under the curve (AUC) of 0.85, 90% sensitivity, and 62% specificity. Adding clinical data slightly improved performance (AUC 0.87, sensitivity 91%, specificity 65%). These findings suggest that ANN combining clinical and imaging data can accurately identify the ischemic core, which may help guide stroke treatment decisions (17).

Role and applications of AIA in vessel analysis in stroke

AIA plays a crucial role in vessel analysis for stroke, particularly in identifying vessel occlusions that guide timely and appropriate treatment decisions (Figure 3). The work of Brugnara et al. demonstrates the power of DL tools in rapidly detecting abnormal blood vessels within minutes, achieving high sensitivity (≥87%) and negative predictive value (≥93%) across multiple hospital settings. Notably, this tool outperformed existing FDA-approved software, highlighting the potential of AI to significantly enhance diagnostic accuracy and speed in AIS management (18).

Figure 3 MIP CTA. The automated occlusion site of the right MCA is marked by a red circle. Cerebral areas with poor collateral flow and perfusion are highlighted in an orange shade. CTA, computed tomography angiography; MCA, middle cerebral artery; MIP, maximum intensity projection.

LVOs represent about 30% of AIS cases and are linked to more severe clinical outcomes (19). In this context, Martinez-Gutierrez et al. evaluated an AI system designed for automatic LVO detection from CT angiographies, which provided real-time alerts to clinicians. The study reported meaningful reductions in critical treatment intervals such as door-to-groin and CT-to-endovascular treatment times by approximately 10 minutes, reflecting improved workflow efficiency (20). Although this did not translate into significant improvements in functional recovery at 90 days, the findings underscore AI’s role in accelerating clinical processes, which is essential for stroke care where every minute counts.

Beyond occlusion detection, AI applications in assessing Collateral Circulation are emerging as important tools in stroke prognosis. Kim et al. developed a DL model that evaluates collateral blood flow using dynamic MR perfusion imaging, showing high accuracy and good agreement with expert ratings. This model could predict infarct growth and patient outcomes, reinforcing the critical impact of Collateral Circulation status and the timing of reperfusion therapy on recovery (21). Since distal occlusions often depend on the adequacy of collateral flow to sustain brain tissue, mapping Collateral Circulation with AI offers valuable prognostic insights. Such analyses can guide clinicians in anticipating treatment success and tailoring reperfusion strategies based on individualized vascular profiles.

Together, these advances illustrate the expanding capabilities of AIA in stroke vessel analysis—from rapid occlusion detection to detailed collateral assessment—enhancing both diagnostic precision and treatment planning. While AI-driven tools have shown clear benefits in reducing treatment delays and improving workflow efficiency, ongoing challenges include translating these gains into long-term functional improvements and integrating Collateral Circulation mapping into routine clinical decision-making. Future research should focus on refining these technologies, validating them across diverse patient populations, and exploring their combined use to optimize stroke management comprehensively.

Role and applications of AIA in MeVOS and DVOs

Despite major advances in AIA for AIS, most AI tools have been developed and validated primarily for LVOs, especially in anterior circulation. In contrast, MeVOs and DVOs are underrepresented, despite their clinical importance and increasing relevance in modern stroke care. MeVOs and DVOs are inherently more challenging to detect due to their location in smaller, more tortuous arterial branches like M2–M4 segments of the MCA or distal PCA branches. These vessels are often poorly visualized on conventional CTA or MRA, making detection difficult even for experienced clinicians. Manual interpretation is subjective, and current AIA systems are not yet optimized to overcome these limitations. The studies by Sriwastwa et al. and Amukotuwa et al. highlight this gap (22,23). While their AI algorithms achieved high sensitivity for proximal LVOs, their performance dropped significantly for MeVOs and posterior circulation strokes. For example, Sriwastwa’s algorithm achieved only 65% sensitivity for MeVOs, and Amukotuwa’s tool, though promising for M2 occlusions, still showed limited specificity and variable sensitivity depending on the site (22,23). These performance limitations likely stem from biases in training data and algorithm design. Most datasets used for model development prioritize larger occlusions, which are easier to annotate and detect. As a result, current AI lacks the fine-grained spatial resolution and pattern recognition capabilities needed to identify small clots or subtle perfusion abnormalities in distal territories. Clinically, this matters. MeVOs and DVOs can still cause significant disability, especially when critical brain regions are affected or when collateral flow is poor. As treatment strategies expand to include these occlusions, the lack of reliable AIA support becomes a bottleneck in care. Improved detection would support faster triage, better selection for endovascular therapy, and more equitable outcomes. Therefore, future research should aim to improve algorithm performance for distal occlusions by incorporating more diverse and representative datasets, refining detection thresholds, and validating performance across varied clinical environments. Without these steps, the full utility of AIA in stroke care -especially in complex or subtle cases- remains out of reach.

Role and applications of AIA in stroke outcome evaluation

AIA has shown growing potential in predicting clinical outcomes in patients with AIS, thus playing an increasingly integral role in post-treatment evaluation and personalized care planning. Particularly in cases involving distal medium vessel occlusions (DMVOs), outcome prediction has traditionally been challenging due to the variability in clinical presentations and therapeutic responses. Karabacak et al. addressed this issue by developing a ML model designed to predict poor outcomes—defined as a modified Rankin Scale (mRS) score of 3–6—in patients with DMVOs. Utilizing a dataset of 164 AIS patients, the model achieved robust predictive performance (AUC =0.815) and identified key prognostic variables, including the National Institutes of Health Stroke Scale (NIHSS) at admission, premorbid mRS, thrombectomy technique, and history of malignancy (24). Notably, the researchers developed an interactive web-based tool for individualized outcome forecasting, underscoring the translational potential of AIA in clinical stroke management. Expanding AI research to encompass a broader range of occlusion types, including DVOs, is vital for establishing comprehensive decision-support systems across all stroke subtypes.

AIA has also been applied to outcome prediction in patients undergoing MT, especially in cases involving LVOs, where early and accurate prognostication can inform both clinical decisions and resource allocation. Von Braun et al. introduced a DL model that integrated pre-treatment CT imaging with clinical variables to predict outcomes in 405 stroke patients treated with endovascular therapy (EVT) (25). The model not only improved the localization of infarcted tissue (Dice score: 0.48–0.52) but also demonstrated lower prediction error for NIHSS scores at discharge (1.5–3.0 points) compared to traditional approaches (25). This framework facilitates precision medicine in stroke care by estimating the extent of brain damage and functional recovery, thereby supporting tailored therapeutic interventions and post-stroke planning.

Post-procedural imaging evaluation is another area where AIA enhances diagnostic accuracy, particularly in distinguishing intracranial hemorrhage (ICH) from contrast staining following EVT—a differentiation that is critical but often difficult using conventional CT imaging. Wang et al. proposed a novel Transformer-based generative adversarial network (Trans-GAN) capable of synthesizing hemorrhage-specific image features from standard single-energy CT scans. Trained on 237 dual-energy CT datasets, the model significantly outperformed existing benchmarks, achieving superior image quality and diagnostic metrics (AUC: 0.88 vs. 0.68; kappa: 0.83 vs. 0.56) (26). By enabling accurate ICH detection in settings lacking advanced imaging modalities, this AI-driven approach improves clinical decision-making and safety monitoring in the post-EVT setting.

AIA has also emerged as a valuable tool for predicting hemorrhagic transformation (HT), a potentially life-threatening complication of reperfusion therapy. Early and reliable prediction of HT is crucial for risk stratification and therapeutic modulation. Several studies have employed DL algorithms that integrate imaging and clinical data to identify patients at heightened risk. Ru et al. developed a weakly supervised DL model that combined NCCT scans with clinical variables to predict HT in patients receiving IVT. This model demonstrated high predictive accuracy, enhancing patient selection and post-treatment vigilance (27). Complementarily, Jiang et al. reported a DL model that utilized pre-treatment imaging and clinical data to accurately forecast HT occurrence, reinforcing the promise of AI for individualized risk assessment (28). Same, Sun et al. developed a semi-automated tool for segmenting hemorrhagic regions on follow-up NCCT scans in patients treated with MT. The algorithm achieved a mean Dice similarity coefficient (DSC) of 0.66, with higher accuracy for parenchymal hematomas (DSC: 0.73) compared to Type 2 hemorrhagic infarctions (DSC: 0.61). Moreover, the tool demonstrated efficient processing (average time: 2.7 seconds per case), indicating potential for real-time prognostic evaluation in clinical workflows (29). These advancements underscore the transformative impact of AIA in stroke outcome evaluation. By enabling early prediction of clinical trajectories, enhancing diagnostic precision, and facilitating personalized care strategies, AIA contributes significantly to optimizing stroke management across the entire continuum of care.

Role and applications of AIA in post-stroke rehabilitation

AI-based imaging analysis is proving to be a valuable tool in post-stroke rehabilitation by enabling dynamic and accurate predictions of patient outcomes. The transformer-based model developed by Klug et al. in 2,492 patients stands out by continuously integrating updated clinical, imaging, and biological data to provide hourly outcome predictions. Its superior performance compared to static models, demonstrated by improved mortality prediction accuracy over the first 72 hours, highlights the advantage of real-time monitoring in stroke care (30). Key predictors such as clinical assessments, treatment timing, and inflammatory markers emphasize the importance of combining diverse data sources for more precise prognostication. This explainable and adaptable approach supports clinicians in making timely, personalized decisions during rehabilitation, potentially improving recovery and optimizing resource use. However, widespread adoption will require validation across different populations and attention to data privacy and integration challenges.

Considering the points mentioned above, AIA plays a significant role with multiple applications in supporting clinical decision-making. AI enhances clinical decision support in stroke by providing personalized treatment recommendations based on imaging data, such as infarct core size, penumbra, and CC status. It helps clinicians decide between IVT or EVT and supports real-time decision-making through rapid neuroimaging analysis. AI also predicts outcomes guiding post-treatment care.

Limitations of AIA

While AI has shown great promise in stroke care, it still has some limitations that need to be addressed. These limitations occur in all aspects of AI use in health care, not just in stroke management. For example, AI algorithms can be biased if the data they are trained on is not diverse or representative of different populations. Vrudhula et al. discuss how bias in healthcare affects patient outcomes, including in medical imaging, where disparities arise from access, acquisition, and interpretation. While ML can enhance diagnostic accuracy and reduce cognitive bias, it may also introduce bias if trained on unbalanced data, poorly designed, or applied inconsistently (31). Ensuring that training datasets are comprehensive and of high quality, it is essential to minimize bias and improve the reliability of AI in clinical settings.

Moreover, while AIA systems can flag abnormalities, clinical expertise remains essential for interpreting these findings within the broader context of the patient’s medical history, symptoms, and overall clinical situation. In a review of Akay et al., 121 studies were analyzed on AI-based clinical decision support systems for AIS, with 65 undergoing full extraction. The study assessed data sources, methodologies, and adherence to AI reporting standards. Findings revealed significant variability in data usage, methodological approaches, and reporting practices, raising concerns about validity and clinical applicability. The review highlights key challenges in translating AI research into clinical practice and provides recommendations to enhance the robustness and implementation of AI-driven decision support in stroke diagnosis and treatment (32).

Additionally, the integration of AIA into clinical workflows presents several challenges. Kotter and Ranschaert highlight barriers to AI integration in radiology, including the lack of standardized implementation protocols, like early PACS adoption challenges (33). They also emphasize the need for radiologists to understand AI both technically and ethically to collaborate effectively with engineers and optimize its use in clinical practice (33). This is not an easily solvable issue, as radiologists need additional training in AI, which requires time and effort. However, integrating AI education into their already demanding clinical practice poses a challenge, as they often have limited time for such training.

Another issue that may be addressed is the fact that AI is probably prone to technical issues in imaging, such as inconsistent image quality and artifacts. As it is good consideration, many efforts have been made to solve this problem. Beljaards et al. developed a DL model to detect and quantify motion artifacts in MRI scans. Trained on both simulated and real data, it identified motion-corrupted scans with up to 96% accuracy and predicted image quality labels with up to 85% accuracy. When used to guide reconstruction based on motion detection, it achieved 93% accuracy, enhancing image quality and minimizing re-acquisition (34). As poor-quality images can lead to inaccurate analysis or missed diagnoses, AIA systems should be better trained to recognize these issues, correct them, and then analyze the images.

Key ethical considerations for AI integration in clinical practice arise to prevent harm and ensure patient safety. For this reason, Word Health Organization (WHO) published a guide on Ethics and Governance of AI for Health as a result among experts in ethics, digital technology, law, human rights, and health ministries. This guide emphasizes the need for ethics and human rights to be central in their design, deployment, and use of AI (35). Central to this guidance is the principle that AI technologies must uphold fundamental human rights and be aligned with core ethical values such as autonomy, beneficence, justice, and non-maleficence. In particular, AI systems must be developed in a manner that avoids reinforcing existing healthcare disparities, ensures informed patient consent, and promotes transparency in decision-making processes. While AI has the potential to significantly improve healthcare delivery, its success and sustainability in clinical practice will depend on a proactive commitment to ethical integrity and human-centered design. Filippi et al. review the risks of bias in AI for Neuroradiology, emphasizing explainability, accountability, and transparency under the principle of “first, do no harm”. They outline steps to reduce bias throughout AI development and discuss fairness criteria to maximize benefits while minimizing harm (36). Regular evaluation and adherence to ethical guidelines are essential for safe clinical integration.

Finally, to the best of our knowledge, research on the cost-effectiveness of AI in stroke care and comparisons between AI systems is lacking. Assessing cost and specificity is crucial to identifying the most efficient AI for clinical use. For instance, one study assessed the cost-effectiveness of AI for detecting LVOs in stroke in the United Kingdom showed that reducing missed LVO diagnoses by 50% at $40 per AI analysis could save $156 per patient. Nationally, this results in $11 million in annual savings (37). The findings highlight AI’s potential to improve stroke care.

Future perspectives of AIA in stroke care

The future of AIA in stroke care appears highly promising, offering the potential to improve both the accessibility and precision of diagnosis and treatment. Two key areas of development include mobile stroke technologies and personalized medicine approaches.

A notable advancement is the integration of AIA into mobile stroke units, which allow for early, prehospital diagnosis and treatment. These systems have been shown to reduce treatment times and improve short-term outcomes. However, their use is mostly limited to large urban hospitals with the necessary infrastructure and financial resources. This creates a gap in access for rural and underserved regions, where the burden of stroke is often higher. While the meta-analysis of more than 21,000 patients supports the effectiveness of mobile stroke units, issues like cost, logistics, and scalability must be addressed before they can be more widely adopted (38). Future efforts should focus on adapting these technologies to settings with fewer resources.

In addition to improving access, AIA is also contributing to more personalized stroke care. AI tools such as the XGBoost model developed by Caliandro et al. show how clinical and imaging data can be combined to predict patient outcomes early in treatment (39). Using variables such as the NIH Stroke Scale (NIHSS), Glasgow Coma Scale (GCS), and imaging findings, these tools help move beyond general treatment protocols toward more tailored care. However, to ensure fairness and reliability, these models must be tested in a wide range of patient populations, including those often underrepresented in clinical studies.

Another important area for future development involves patients who do not respond to standard treatments like IVT or EVT, especially in MeVOS and DVOS (7,8). AI has the potential to help identify which patients are more or less likely to benefit from specific treatments, based on factors such as vessel anatomy, collateral circulation, etc. However, combining such detailed information into clinical decision-making brings new challenges, such as handling complex data, ensuring compatibility across systems, and making AI outputs easy for clinicians to interpret.

In summary, AIA offers great opportunities to improve stroke care by increasing access and supporting more personalized treatment decisions. To fully realize this potential, challenges such as limited access, data diversity, ethical use, and clinical adoption must be addressed. Building trust in AI systems and ensuring they are practical and scalable in real-world settings will be essential for their successful integration into everyday stroke care.

Limitations of this review

This review has some limitations. The literature search was limited to studies published between January 2019 and January 2025. While this timeframe was chosen to focus on the most recent advances, it could limit the historical context or overlook nascent developments. Although multiple databases were used, unpublished studies, conference papers, and non-English publications might have been missed, potentially introducing publication and language bias. Additionally, although efforts were made to include a wide range of clinical scenarios, limited data are currently available on AIA applications in MeVOs and DMVOs. As a result, we were unable to offer an extensive or detailed discussion of this important and growing area. As a narrative review, formal quality assessment and meta-analysis were not performed, limiting the ability to quantitatively synthesize findings. The wide variety of AI applications, methodologies, and patient populations across studies makes direct comparisons difficult and may affect the generalizability of conclusions. Another limitation stems from potential publication bias, as positive or novel AI applications are more likely to be published, while negative or inconclusive findings may be underreported. This could lead to an overly optimistic view of AI’s current capabilities. Lastly, the rapid pace of AI development means that some information may quickly become outdated, and real-world validation and evolving ethical considerations require ongoing attention.

Summary

This review highlights the significant potential of AIA in improving stroke care, particularly in the rapid detection of ischemic lesions and gives highlights into distal occlusions. AI-driven tools have demonstrated impressive accuracy in stroke diagnosis and have helped reduce treatment delays, thus enhancing clinical decision-making. The integration of these technologies into clinical practice could significantly impact patient outcomes by enabling faster and more precise interventions. To summarize the key AI applications in stroke care, Table 2 provides an overview of studies and their findings.

Conclusions

AIA is transforming stroke care by enhancing diagnostic accuracy, guiding treatment decisions, and predicting outcomes. However, its limited performance in detecting medium and distal occlusions, along with challenges in clinical integration and ethical considerations, highlights the need for continued research and refinement to ensure broader, reliable, and equitable application in clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jni.amegroups.com/article/view/10.21037/jni-25-15/rc

Peer Review File: Available at https://jni.amegroups.com/article/view/10.21037/jni-25-15/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jni.amegroups.com/article/view/10.21037/jni-25-15/coif). P.P. serves as an unpaid editorial board member of Journal of Neurointervention from November 2024 to December 2026. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Cui S, Tseng HH, Pakela J, et al. Introduction to machine and deep learning for medical physicists. Med Phys 2020;47:e127-47. [Crossref] [PubMed]
2. Oren O, Gersh BJ, Bhatt DL. Artificial intelligence in medical imaging: switching from radiographic pathological data to clinically meaningful endpoints. Lancet Digit Health 2020;2:e486-8. [Crossref] [PubMed]
3. Global, regional, and national burden of stroke and its risk factors, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet Neurol 2024;23:973-1003. [Crossref] [PubMed]
4. Martin SS, Aday AW, Allen NB, et al. 2025 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. Circulation 2025;151:e41-e660. [Crossref] [PubMed]
5. Berge E, Whiteley W, Audebert H, et al. European Stroke Organisation (ESO) guidelines on intravenous thrombolysis for acute ischaemic stroke. Eur Stroke J 2021;6:I-LXII. [Crossref] [PubMed]
6. Turc G, Bhogal P, Fischer U, et al. European Stroke Organisation (ESO) - European Society for Minimally Invasive Neurological Therapy (ESMINT) Guidelines on Mechanical Thrombectomy in Acute Ischemic Stroke. J Neurointerv Surg 2023;15:e8. [Crossref] [PubMed]
7. Goyal M, Ospel JM, Ganesh A, et al. Endovascular Treatment of Stroke Due to Medium-Vessel Occlusion. N Engl J Med 2025;392:1385-95. [Crossref] [PubMed]
8. Psychogios M, Brehm A, Ribo M, et al. Endovascular Treatment for Stroke Due to Occlusion of Medium or Distal Vessels. N Engl J Med 2025;392:1374-84. [Crossref] [PubMed]
9. Shafaat O, Sotoudeh H. Stroke Imaging. [Updated 2023 May 1]. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2025 Jan-. Available online: https://www.ncbi.nlm.nih.gov/books/NBK546635/
10. Soun JE, Chow DS, Nagamine M, et al. Artificial Intelligence and Acute Stroke Imaging. AJNR Am J Neuroradiol 2021;42:2-11. [Crossref] [PubMed]
11. Adamou A, Beltsios ET, Bania A, et al. Artificial intelligence-driven ASPECTS for the detection of early stroke changes in non-contrast CT: a systematic review and meta-analysis. J Neurointerv Surg 2023;15:e298-304. [Crossref] [PubMed]
12. Qiu W, Kuang H, Teleg E, et al. Machine Learning for Detecting Early Infarction in Acute Stroke with Non-Contrast-enhanced CT. Radiology 2020;294:638-44. [Crossref] [PubMed]
13. Öman O, Mäkelä T, Salli E, et al. 3D convolutional neural networks applied to CT angiography in the detection of acute ischemic stroke. Eur Radiol Exp 2019;3:8. [Crossref] [PubMed]
14. Gheibi Y, Shirini K, Razavi SN, et al. CNN-Res: deep learning framework for segmentation of acute ischemic stroke lesions on multimodal MRI images. BMC Med Inform Decis Mak 2023;23:192. [Crossref] [PubMed]
15. Nicolas-Jilwan M, Wintermark M. Automated Brain Perfusion Imaging in Acute Ischemic Stroke: Interpretation Pearls and Pitfalls. Stroke 2021;52:3728-38. [Crossref] [PubMed]
16. Fainardi E, Busto G, Morotti A. Automated advanced imaging in acute ischemic stroke. Certainties and uncertainties. Eur J Radiol Open 2023;11:100524. [Crossref] [PubMed]
17. Kasasbeh AS, Christensen S, Parsons MW, et al. Artificial Neural Network Computer Tomography Perfusion Prediction of Ischemic Core. Stroke 2019;50:1578-81. [Crossref] [PubMed]
18. Brugnara G, Baumgartner M, Scholze ED, et al. Deep-learning based detection of vessel occlusions on CT-angiography in patients with suspected acute ischemic stroke. Nat Commun 2023;14:4938. [Crossref] [PubMed]
19. Lakomkin N, Dhamoon M, Carroll K, et al. Prevalence of large vessel occlusion in patients presenting with acute ischemic stroke: a 10-year systematic review of the literature. J Neurointerv Surg 2019;11:241-5. [Crossref] [PubMed]
20. Martinez-Gutierrez JC, Kim Y, Salazar-Marioni S, et al. Automated Large Vessel Occlusion Detection Software and Thrombectomy Treatment Times: A Cluster Randomized Clinical Trial. JAMA Neurol 2023;80:1182-90. [Crossref] [PubMed]
21. Kim YC, Chung JW, Bang OY, et al. A Deep Learning-Based Automatic Collateral Assessment in Patients with Acute Ischemic Stroke. Transl Stroke Res 2023;14:66-72. [Crossref] [PubMed]
22. Sriwastwa A, Aziz YN, Weiss K, et al. Performance of Automated Algorithm in Large and Medium Vessel Occlusion Detection: A Real-World Experience. AJNR Am J Neuroradiol 2025;46:476-82. [Crossref] [PubMed]
23. Amukotuwa SA, Straka M, Dehkharghani S, et al. Fast Automatic Detection of Large Vessel Occlusions on CT Angiography. Stroke 2019;50:3431-8. [Crossref] [PubMed]
24. Karabacak M, Ozkara BB, Faizy TD, et al. Data-Driven Prognostication in Distal Medium Vessel Occlusions Using Explainable Machine Learning. AJNR Am J Neuroradiol 2025;46:725-32. [Crossref] [PubMed]
25. von Braun MS, Starke K, Peter L, et al. Prediction of tissue and clinical thrombectomy outcome in acute ischaemic stroke using deep learning. Brain 2025;148:2348-60. [Crossref] [PubMed]
26. Wang T, Jiang C, Ding W, et al. Deep-Learning Generated Synthetic Material Decomposition Images Based on Single-Energy CT to Differentiate Intracranial Hemorrhage and Contrast Staining Within 24 Hours After Endovascular Thrombectomy. CNS Neurosci Ther 2025;31:e70235. [Crossref] [PubMed]
27. Ru X, Zhao S, Chen W, et al. A weakly supervised deep learning model integrating noncontrasted computed tomography images and clinical factors facilitates haemorrhagic transformation prediction after intravenous thrombolysis in acute ischaemic stroke patients. Biomed Eng Online 2023;22:129. [Crossref] [PubMed]
28. Jiang L, Zhou L, Yong W, et al. A deep learning-based model for prediction of hemorrhagic transformation after stroke. Brain Pathol 2023;33:e13023. [Crossref] [PubMed]
29. Sun J, Werdiger F, Blair C, et al. Automatic segmentation of hemorrhagic transformation on follow-up non-contrast CT after acute ischemic stroke. Front Neuroinform 2024;18:1382630. [Crossref] [PubMed]
30. Klug J, Leclerc G, Dirren E, et al. Machine learning for early dynamic prediction of functional outcome after stroke. Commun Med (Lond) 2024;4:232. [Crossref] [PubMed]
31. Vrudhula A, Kwan AC, Ouyang D, et al. Machine Learning and Bias in Medical Imaging: Opportunities and Challenges. Circ Cardiovasc Imaging 2024;17:e015495. [Crossref] [PubMed]
32. Akay EMZ, Hilbert A, Carlisle BG, et al. Artificial Intelligence for Clinical Decision Support in Acute Ischemic Stroke: A Systematic Review. Stroke 2023;54:1505-16. [Crossref] [PubMed]
33. Kotter E, Ranschaert E. Challenges and solutions for introducing artificial intelligence (AI) in daily clinical workflow. Eur Radiol 2021;31:5-7. [Crossref] [PubMed]
34. Beljaards L, Pezzotti N, Rao C, et al. AI-based motion artifact severity estimation in undersampled MRI allowing for selection of appropriate reconstruction models. Med Phys 2024;51:3555-65. [Crossref] [PubMed]
35. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Executive Summary. 28 June 2021. ISBN: 9789240029200. Available online: https://www.who.int/publications/i/item/9789240037403
36. Filippi CG, Stein JM, Wang Z, et al. Ethical Considerations and Fairness in the Use of Artificial Intelligence for Neuroradiology. AJNR Am J Neuroradiol 2023;44:1242-8. [Crossref] [PubMed]
37. van Leeuwen KG, Meijer FJA, Schalekamp S, et al. Cost-effectiveness of artificial intelligence aided vessel occlusion detection in acute stroke: an early health technology assessment. Insights Imaging 2021;12:133. [Crossref] [PubMed]
38. Fatima N, Saqqur M, Hussain MS, et al. Mobile stroke unit versus standard medical care in the management of patients with acute stroke: A systematic review and meta-analysis. Int J Stroke 2020;15:595-608. [Crossref] [PubMed]
39. Caliandro P, Lenkowicz J, Reale G, et al. Artificial intelligence to predict individualized outcome of acute ischemic stroke patients: The SIBILLA project. Eur Stroke J 2024;9:1053-62. [Crossref] [PubMed]