The advent of artificial intelligence (AI) has profoundly transformed numerous disciplines, including the critical field of healthcare. AI has rapidly emerged as a transformative and beneficial tool in the crucial diagnosis and comprehensive management of acute stroke. Stroke is a health problem that is of significant concern. Stroke is divided into two primary types: ischemic and hemorrhagic. It is the fifth leading cause of death in the United States and a major cause of disability worldwide [1]. Each year, approximately 800,000 individuals in the USA face the devastating reality of a new or recurrent stroke, underscoring the urgent need for rapid, accurate, and effective diagnostic procedures to significantly improve clinical outcomes for patients [2, 3]. The concept of the “Golden Hour” is a term used to emphasize the critical and time-sensitive nature of timely intervention. It is during this limited time period that various treatment options can most effectively mitigate and alleviate the enduring consequences of a stroke [4, 5]. Nevertheless, the recognition of acute stroke symptoms in emergency medical settings presents a considerable and complex challenge. This is due to the intricacies and time-sensitive demands of emergency care. Consequently, delays in treatment can result in the loss of opportunities for effective and timely interventions that could ultimately save lives and improve the quality of recovery for stroke patients [1, 6].

Multimodal imaging approaches—including non-contrast Computed Tomography (CT), CT angiography, CT perfusion, and the diffusion-weighted imaging (DWI)–fluid-attenuated inversion recovery (FLAIR) mismatch—are central to contemporary acute ischemic stroke evaluation and management. Among these modalities, DWI has a distinctive role, as it represents the most sensitive magnetic resonance imaging (MRI) sequence for the detection of acute ischemic infarction once MRI is performed. Consequently, DWI is routinely incorporated into the diagnostic workflow of acute ischemic stroke and serves as a critical imaging modality in clinical practice [7]. Inadequate or delayed diagnosis can lead to poorer clinical outcomes, emphasizing the necessity for robust clinical support systems in emergency departments [8]. The ongoing Fourth Industrial Revolution has furthered the application of AI technologies, particularly in the field of medical imaging [9]. The incorporation of AI into stroke diagnosis has been demonstrated to enhance diagnostic efficiency and hold the potential to improve patient survival rates and functional recovery outcomes [10, 11, 12]. Therefore, the deployment of AI in acute stroke diagnosis is a promising frontier in the pursuit of more effective and timely patient care in emergency departments with limited expertise in stroke management [13]. While emergency MRI availability is not universal, many comprehensive and tertiary stroke centers routinely incorporate MRI-DWI into their acute stroke protocols, especially when diagnostic uncertainty exists or when stroke mimics must be excluded. In this context, an AI-based system trained exclusively on DWI may provide robust lesion detection and serve as a valuable decision-support tool in MRI-equipped emergency departments rather than as a replacement for multimodal imaging strategies.

The progressive development of AI in the field of stroke diagnosis signifies a pivotal advancement in medical technology, with the potential to enhance diagnostic efficiency and contribute to the preservation of lives through early intervention strategies. In order to achieve a desirable level of generalizability for any AI system, it is necessary to validate it using external databases with different clinical settings, imaging protocols, and populations. Additionally, a comparative analysis with radiologists is necessary to ascertain the relative efficacy of these systems in actual clinical settings [12, 14]. In the present study, the validity of a DWI-MRI-based deep learning model used for acute stroke detection is evaluated. In addition, the performance of the model was assessed through the analysis of databases from multiple centers, with the goal of determining its efficacy in routine clinical practice.

This retrospective study investigates the results of the application of an AI system based on diffusion MRI findings in acute ischemic stroke and non-stroke patients. The model was developed and validated both internally and externally.

2.3 Artificial Intelligence Study

The pre-processed diffusion MRI images were used as input to the deep learning-based model for differentiation of acute ischemic stroke and normal cases. In the diagnostic process, image features are automatically extracted using a deep learning algorithm, and the most relevant features are identified for the classification task. The model was trained using slice-level DWI images. However, all dataset splitting was performed at the patient level to ensure independence, and no slices from the same patient were included in more than one dataset. For performance evaluation, slice-level predictions were aggregated at the patient level using a majority-voting approach, whereby a patient was classified as acute ischemic stroke if the majority of slices were predicted as positive. The model simultaneously performs the classification process based on the best and informative deep features. In deep learning, image features are extracted and analyzed hierarchically from raw data to high-level concepts, in different layers of the network. The architecture of the proposed convolutional neural network (CNN) is provided in Fig. 1. MRI images were resized to 140 $\times$ 140 pixels and were fed into the model. The model consists of two convolutional and seven residual blocks. The convolutional block (ConvBlock) starts with a convolutional layer followed by a batch normalization (BatchN) and a rectified linear unit (ReLU) layer. The output of the first ConvBlock is split into two branches in the residual block (ResBlock). The first branch consists of a convolution layer followed by a BatchN layer, a ReLU activation, a convolutional layer, and a BatchN layer. Then, the outputs of this branch and the skip connection are added using an addition layer. On the other hand, the last two ResBlocks consist of a convolution layer followed by BatchN and ReLU layers with a skip connection. Between all ResBlocks, a pooling layer is put to reduce the dimension of data and preserve meaningful features. The size of filters in all convolution layers is 3 $\times$ 3. At the end, a global max (GlobMax) layer followed by two dense and dropout layers is used to prevent overfitting (Fig. 1).

Fig. 1.

The proposed deep learning model used for stroke detection.

In addition to dropout layer, an augmentation technique was used to prevent overfitting and increase generalizability. In this study, MRI images were randomly rotated (–30, 30 degrees) and scaled (0.6, 1). The model was trained using stochastic gradient descent (sgdm) optimizer, cross-entropy loss function, initial learning rate of 0.001, validation frequency of 2, and L2 regularization of 0.001. Before each training epoch, the training dataset was shuffled to prevent learning order-specific patterns and improve convergence of the model. Hyperparameters were determined through preliminary empirical testing aimed at achieving stable training behavior and preventing overfitting. A formal hyperparameter optimization procedure (e.g., grid search or automated tuning) was not performed in this study.

The classification threshold was defined according to the training set using Receiver Operating Characteristic (ROC) analysis and subsequently fixed. This same threshold was applied without modification to the validation datasets when calculating sensitivity, specificity, and accuracy, ensuring consistent and unbiased performance evaluation.

The dataset from center 1 was split into 85% for training and tuning the model and 15% for the validation phase. Data from centers 2 and 3 were considered for the external validation step. In addition to deep learning, the internal and external validation datasets were reviewed by three expert radiologists in neuroimaging with 7 years (center 1), 5 years (center 2), and 7 years (center 3) to diagnose acute ischemic stroke cases. No cross-validation or permutation testing was performed in this study.

In this multicenter study, a DWI-MRI-based deep learning model demonstrated high diagnostic accuracy for acute ischemic stroke, as evidenced by the model’s performance on both internal and external validation datasets. The model demonstrated a 100% accuracy rate in the internal validation and up to a 99% accuracy rate in external centers, exhibiting performance comparable to that of radiologists. The findings indicate that AI-based decision support systems can play a substantial role in facilitating rapid and reliable diagnosis, particularly within emergency departments where specialist neuroradiology support is scarce or in high-volume settings. Although perfect performance in internal validation may raise concerns regarding overfitting, the consistent results observed across two independent external validation cohorts support the robustness and generalizability of the proposed model.

Recent studies have demonstrated that AI has been shown to yield results similar to those of radiologists in stroke imaging (non-contrast CT, CT angiography, CT perfusion, MR angiography, DWI) for lesion detection, Alberta stroke program early CT score (ASPECTS) scoring, and detection of large vessel occlusions [15, 16]. The development of deep learning models tailored for DWI has yielded high levels of sensitivity and specificity, achieved by minimizing error rates in the detection of small lesions. As stated by Liu et al. [16], the utilization of deep learning algorithms on DWI data yielded diagnostic accuracies that were comparable to those attained by radiologists with extensive experience. Similarly, Abedi et al. [1] underscored the potential of AI-supported systems to expedite emergency department workflow and reduce mortality. The present study lends support to these findings in the existing literature. The multicenter design of the study serves to strengthen the model’s generalizability. The high performance demonstrated, even with data obtained from different devices and protocols, highlights the effectiveness of the preprocessing (histogram matching, normalization) and data augmentation methods used.

Accurate and prompt diagnosis of acute stroke is paramount for effective utilization of time windows for intravenous thrombolysis and mechanical thrombectomy. AI-supported automatic analyses have the potential to enhance safety measures, especially during the triage and diagnostic processes within emergency departments. The existing literature indicates that the implementation of automated software, such as RAPID, has been demonstrated to enhance clinical outcomes [17, 18, 19]. From a practical perspective, several challenges must be addressed before clinical deployment, including seamless integration with PACS, regulatory approval, and ensuring real-time or near real-time processing in emergency settings. Additionally, robust validation across diverse imaging protocols and scanners is essential to support safe and reliable implementation.

It should be emphasized that the proposed model is not intended to replace multimodal stroke imaging workflows but rather to complement existing diagnostic pathways in MRI-equipped centers. The reliance on DWI alone may limit applicability in emergency settings where MRI access is restricted; however, in centers with rapid MRI capability, automated DWI-based stroke detection may enhance diagnostic confidence, reduce interpretation time, and support clinical decision-making, particularly in high-volume or resource-limited environments.

The Gradient-weighted Class Activation Mapping (Grad-CAM) method, as implemented in our study, demonstrated that the model focused on lesion areas when making decisions. Grad-CAM is a visualization technique that generates class-specific activation maps using the gradient information in the last convolutional layer to explain the decision-making processes of deep convolutional neural networks (CNNs). In CNN models trained for ischemic stroke detection in brain MRI images, the Grad-CAM method enables insightful analysis by localizing the spatial regions that guide the model’s prediction and offers the opportunity to assess the clinical relevance of the decision mechanism. Specifically, when contrasting ischemic lesions and normal tissues in diffusion-weighted images, heat maps derived from Grad-CAM can directly illustrate the location and extent of the lesion. In a study by Lai et al. [20], Grad-CAM outputs, which allow for visual assessment of the location and extent of lesions in diffusion-weighted images, are used as an important tool to evaluate whether the model correctly processes pathological regions. Other recent studies in the literature have also shown that Grad-CAM improves both model performance and explainability in automatic ischemic stroke detection systems [21, 22]. Consequently, Grad-CAM not only fosters classification accuracy but also assumes a pivotal role in the implementation of AI-based clinical decision support systems. In a small number of cases, Grad-CAM visualizations highlighted regions adjacent to, rather than directly overlapping with, radiologists’ annotated lesions. Such discrepancies may reflect sensitivity to subtle diffusion changes, partial volume effects, or image noise, and underscore the complementary role of AI-based attention maps rather than direct lesion delineation.

The retrospective design of this study may introduce selection bias. Another major limitation is the exclusion of patients with prior stroke and lacunar infarctions, which may limit real-world applicability. Chronic diffusion abnormalities following previous stroke can overlap with acute lesions, complicating reliable labeling, while the slice-level classification approach used in this study is less sensitive to very small lacunar infarcts. In addition, stroke subtypes and the topographic distribution of ischemic lesions (e.g., vascular territories or lobar location) were not systematically analyzed, as etiological and anatomical localization were beyond the scope of this detection-focused study. Accordingly, the reported diagnostic performance should be interpreted as the model’s ability to detect DWI-visible diffusion abnormalities rather than as a fully comprehensive clinical diagnosis of acute ischemic stroke. The exclusive use of DWI without integration of other imaging modalities such as CTA, CTP, or DWI–FLAIR mismatch may restrict generalizability to centers without emergency MRI availability. Future studies should incorporate multimodal imaging data, lesion-level segmentation, and prospective designs to evaluate workflow impact and clinical outcomes, as well as broader international external validation to enhance robustness and clinical applicability.

In summary, the present study demonstrated that the DWI MRI-based deep learning model provides diagnostic accuracy comparable to that of radiologists in the diagnosis of acute ischemic stroke and can be reliably applied in multicenter settings. In the future, the integration of AI systems into the clinical workflow for acute stroke care may be facilitated by prospective and multimodal studies.

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

TYK and AAA designed the study. BNK, MEA, SY, MD collected the data. AAA, MSÇ, AM, TYK analyzed the data and prepared the figures. BNK, MEA, SY, MD and MSÇ drafted the manuscript. TYK, AAA, AM brought major revisions in significant proportions of the manuscript. TYK and AM supervised the study, provided important intellectual input, and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

The study was conducted in accordance with the Declaration of Helsinki. Ethical approval was granted by the Institutional Review Board of the University of Health Sciences, Sancaktepe Şehit Prof. Dr. İlhan Varank Training and Research Hospital, Istanbul, Türkiye (Approval No: 2023-91). Due to the retrospective nature of the study, the requirement for informed consent was waived.

Not applicable.

This research received no external funding.

The authors declare no conflict of interest.