Acute ischemic stroke (AIS) is a time-critical neurological emergency caused by interruption of cerebral blood flow, resulting in irreversible neuronal injury if timely reperfusion is not achieved.[1] Stroke continues to be a leading cause of death and disability worldwide.[2] The Global Burden of Disease Study 2021 estimated that, at a global level, 11.9 million new stroke events, 93.8 million prevalent cases of stroke, and 7.3 million deaths were caused in 2021, representing 160.5 million disability-adjusted life-years (DALYs).[3] The largest proportion of incident strokes was ischemic stroke, accounting for about 65.3% of all incident strokes worldwide.[4]

The key to diagnosis and management of AIS is accurate and timely neuroimaging to guide treatment, including intravenous thrombolysis and endovascular thrombectomy, which depend on the presence, degree, location, and timing of cerebral ischemia.[5] Magnetic resonance imaging (MRI), especially diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) measurement, is highly sensitive for the detection of acute ischemic lesions and can detect small or early infarcts that might not be appreciated on conventional imaging.[6] Modern stroke treatment guidelines also call for expedient vascular and tissue imaging if needed for treatment selection, especially in those patients who present outside of traditional treatment windows.[7]

Although it is a valuable tool for diagnosis, the conventional MRI interpretation relies on the knowledge, volume of cases, and experience of the reporting radiologist.[8] Subtle cortical infarcts, small lacunar lesions, posterior circulation strokes, and hyperacute lesions can be difficult to diagnose, leading to false-negative interpretation or late diagnosis.[9] These restrictions have led to growing interest in the use of artificial intelligence (AI) for DWI and ADC image analysis, identification of ischemic abnormalities, and suspicious area demarcation for radiologist review.[10] The pooled sensitivity and specificity for the use of MRI-based AI studies in detecting ischemic stroke were estimated at 93% and 93%, respectively, in a systematic review and meta-analysis, with significant differences in study design, algorithm validation, and clinical utility revealed.[11]

Recent evidence indicates that AI can offer value beyond automated image classification, as it enhances human interpretation. In a multi-reader study of 407 MRI examinations, an MRI-based deep-learning algorithm improved clinicians' sensitivity from 77% to 88% and area under the receiver operating characteristic curve (AUC) from 0.90 to 0.93, while also improving inter-reader agreement.[12] A similarly recent study on challenging AIS lesions showed that pooled sensitivity had improved from 74.6% to 90.6%, and AUC from 0.85 to 0.93, with AI assistance, especially in the detection of small and hyperacute lesions.[13]  In addition, a multicenter, external validation study from 2026 showed that a DWI-based deep-learning model achieved 100% sensitivity and 98% specificity in independent external validation datasets, and comparable diagnostic performance to expert radiologists.[14] The results suggest that AI has the potential to support radiological assessment in the diagnosis of the disease and does not replace it completely.

However, the key is not whether AI can detect ischemic changes on its own, but whether AI-supported MRI offers a quantifiable diagnostic benefit in the clinical setting for the detection of acute ischemic stroke. The diagnostic accuracy might vary significantly between different AI algorithms, readers, and lesions, and excellent results on controlled AI datasets do not necessarily indicate the same performance in routine clinical practice. Therefore, it is clinically relevant to compare the AI-assisted MRI with conventional radiological assessment to determine if AI can enhance sensitivity, specificity, overall accuracy, and diagnostic discrimination without replacing the role of expert radiological assessment. The present study was therefore designed to comparatively assess the diagnostic accuracy of artificial intelligence-assisted MRI and conventional radiological assessment of acute ischemic stroke.

The Comparative diagnostic accuracy study was conducted in the Department of Radiology, Mayo Hospital, Lahore, for six months between 1st January 2026 and 30th June 2026. The sample size was calculated using the OpenEpi software based on an expected sensitivity of 89%, a 95% confidence level, and a 10% absolute margin of error. The minimum number of participants required with acute ischemic stroke was approximately 38. Based on an anticipated prevalence of 25% in the population being tested for stroke on MR images, the number of participants needed was calculated as 152.[15] After allowing 10% for incomplete or non-diagnostic examinations, the final sample size was increased to 168 participants. A consecutive non-probability sampling technique was used. Patients aged 18 years or older, with acute neurological symptoms suggestive of ischemic stroke and who had brain MRI during the study period, were included. Patients were included if either the MRI examination contained diagnostic-quality DWI and ADC sequences or if clinical data were available to make the final diagnosis. All patients with confirmed acute ischemic stroke and all patients whose clinical presentation was finally deemed not to be ischemic were included in the analysis to allow assessment of the diagnostic value of both assessment methods in an appropriate clinical population. Patients with known chronic neurological lesions that significantly affected the interpretation of the MRI were excluded, as were patients with technically substandard or severely motion-degraded MRI examinations, and examinations lacking or non-diagnostic DWI or ADC sequences. Major intracranial hemorrhage, presence of brain tumors, significant traumatic brain injury, or other structural abnormalities in which it was impossible to successfully assess acute ischemia were also excluded. Incomplete clinical or imaging data prevented a definitive diagnosis in these patients, and they were not included in the final analysis. After institutional ethical approval, eligible patients were enrolled consecutively after obtaining informed consent. The demographic and clinical data, such as the type of symptoms, age, sex, time duration from onset of symptoms to MRI, relevant vascular risk factors, and previous history of stroke, hypertension, diabetes mellitus, and AF, were collected on a structured data-collection proforma. Initial neurological assessment and laboratory and clinical data were also recorded, if available. All brain MRI scans were obtained using the same stroke imaging protocol used at the hospital. Images of DWI and ADC were analyzed for acute ischemic lesions, along with the routine sequences acquired, such as FLAIR. The MRI was assessed independently by a qualified radiologist who did not know the AI result for the conventional assessment. Acute ischemic stroke presence, lesion location, lesion number, and other imaging qualities were documented by the radiologist. The MRI data was subsequently preprocessed and fed into the specific AI stroke detection model for an AI-assisted assessment. Acute ischemic lesions were suspected, and the presence or absence of the lesions, as well as, if available, the location or the probability score provided by the algorithm, were recorded separately from the AI output. The AI assessment was performed without modification of the original MRI images. The conventional radiologist's interpretation and AI output were kept separate during the initial assessment to avoid incorporation bias. A final reference diagnosis was subsequently established by consensus of experienced neuroradiologists, with consideration of the complete MRI examination and relevant clinical information. The reference diagnosis used for calculating the diagnostic accuracy was acute ischemic stroke present or absent. The collected data were entered and analyzed using SPSS version 27.0. Continuous variables like age and time from onset of symptoms to MRI were tested for normality with the Shapiro–Wilk test and then presented as mean ± SD and as median (IQR). Categorical variables were displayed as frequencies and percentages. The accuracy of AI-enhanced MRI compared to the gold standard diagnostic finding was tested. The sensitivity, specificity, PPV, NPV, diagnostic accuracy, positive likelihood ratio, negative likelihood ratio, and 95% confidence intervals were computed. ROC curves were plotted, and the area under the ROC curve (AUROC) was computed for both methods. The AUROCs of AI-assisted MRI and conventional radiological assessment were compared by using the DeLong test for correlated ROC curves. Cohen's kappa coefficient was used to assess agreement between the AI assessment and reference diagnosis and between the conventional radiological assessment and reference diagnosis. McNemar's tests were applied for paired categorical diagnostic outcomes to compare the proportions of cases correctly classified based on AI-assisted MRI with the proportions of cases correctly classified based on conventional radiological assessment. Subgroup analysis was conducted in the following subgroups when adequate numbers of observations were present: age, sex, time from symptom onset to MRI, lesion size, and anatomical location. A two-sided p-value <0.05 was considered statistically significant.

The study included 168 participants with a mean age of 58.7 ± 13.2 years; 101 (60.1%) were male, and 67 (39.9%) were female. The most common vascular risk factors were hypertension (64.3%), dyslipidaemia (42.3%), and diabetes mellitus (39.9%). The median time interval between symptom onset and MRI was 7.0 hours (IQR 4.0–12.0), of which 72/171 (42.9%) had MRI within 6 hours of symptom onset. (Table 1)

The final reference diagnosis was acute ischemic stroke (AIS) in 42 (25.0%) participants and no acute ischemic stroke in 126 (75.0%). The most frequent clinical manifestations were focal weakness, facial asymmetry, and speech disturbance. AIS patients who were confirmed had a median NIHSS of 8 (IQR 5–13). (Table 2)

In 42 participants with AIS, DWI showed acute infarction in all; ADC restriction in 40 (95.2%). There was a single ischemic lesion in most (69.0%) patients, and the most common size category of lesion was 10-30 mm (42.9%). Anterior circulation involvement predominated (81.0%), and half of the AIS cases had involvement of the cortex. (Table 3)

The AI-assisted MRI correctly identified 39 out of 42 AIS cases and classified 118 out of 126 non-AIS cases as negative, with a sensitivity of 92.9% and specificity of 93.7%. The overall diagnostic accuracy of the test was 93.5%, with a positive predictive value of 83.0% and a negative predictive value of 97.5%. The positive and negative likelihood ratios were 14.64 and 0.08, respectively. (Table 4)

Conventional radiological assessment identified 33 of 42 AIS cases and correctly classified 112 of 126 participants without AIS. Its sensitivity was 78.6%, and specificity was 88.9%; positive predictive value was 70.2% and negative predictive value was 92.6%. The overall diagnostic accuracy was 86.3%, the positive likelihood ratio was 7.07, and the negative likelihood ratio was 0.24. (Table 5)

AI-assisted MRI was more sensitive, specific, PPV, NPV, and diagnostic accuracy when compared to conventional radiological assessment. The absolute differences were: 14.3 percentage points for sensitivity, 4.8 for specificity, 12.8 for PPV, 4.9 for NPV, and 7.2 for accuracy. The result of McNemar's test suggests there is a statistically significant difference between paired diagnostic classifications (p=0.013). (Table 6)

The AUROC for AI-assisted MRI was 0.933 (95% CI: 0.884–0.982), and that for conventional radiological assessment was 0.851 (95% CI: 0.780–0.922), as determined by ROC analysis. The difference between the two AUROCs was 0.082 (95% CI: 0.021–0.143) and was statistically significant on DeLong testing (p=0.009). (Table 7)

Table 1. Demographic and baseline characteristics of study participants (n=168)

Table 2. Clinical presentation and final reference diagnosis (n=168)

Table 3. MRI characteristics among patients with acute ischemic stroke (n=42)

Table 4. Diagnostic classification by AI-assisted MRI compared with the reference diagnosis

Table 5. Diagnostic classification by conventional radiological assessment compared

with the reference diagnosis

Table 6. Comparison of diagnostic performance between AI-assisted MRI and

conventional radiological assessment

Table 7. Receiver operating characteristic analysis of AI-assisted MRI and conventional radiological assessment

The present study showed that AI-assisted MRI had high diagnostic accuracy in detecting acute ischemic stroke, and exhibited excellent sensitivity of 92.9%, specificity of 93.7%, and overall accuracy of 93.5%. These findings are consistent with the growing evidence that artificial intelligence can reliably identify acute ischemic lesions on brain MRI. The pooled sensitivity and specificity values for the 2024 systematic review and meta-analysis of 33 studies were around 93% and 93%, respectively, very similar to the values seen in the present study.[11] In an external test of a 3D deep convolutional neural network developed from multiparametric brain MRI, Nael et al. (2021) found sensitivity of 90% and specificity of 97% and an AUROC of 0.97 for identifying acute infarction. The overall diagnostic discrimination was similar, although the AI sensitivity was somewhat higher and specificity somewhat lower. Differences may relate to the imaging sequences, datasets, prevalence of AIS, and external validation populations used in the two studies.[16] Liu et al. (2021) evaluated deep learning for detection and segmentation of diffusion abnormalities using 2,348 clinical DWI examinations and an external dataset of 280 MRIs. They presented good performance in the case of small lesions and remained robust under the influence of artefacts, low resolution, and technical heterogeneity. This is important for the present study, since DWI-based MRI may help in the assessment of small, possibly subtle ischemic lesions, which are a significant diagnostic challenge for conventional visual interpretation.[17] Bridge et al. (2022) reported 92% accuracy, 96% sensitivity, 87% specificity, and an AUROC of 0.98 for a deep learning model to detect acute infarction from MRI. That study reported a sensitivity of 92.9%, slightly lower than the 94.4% seen here, and a specificity of 90.4%, which is lower than the 92.6% seen here. The similarly high AUROC values indicate that both approaches provided strong discrimination between ischemic and non-ischemic examinations.[18] Krag et al. (2023) compared a commercially available deep learning algorithm in 995 suspected stroke patients and found its sensitivity and specificity for acute ischemic lesions to be 89% and 90%, respectively. In the present study, there was slight improvement in the sensitivity and specificity. Importantly, their study revealed that the characteristics of the lesion and imaging quality played a role in performance, with DWI artifacts decreasing specificity and larger lesions increasing sensitivity. These observations are pertinent to the present results because the extent of the lesion and the quality of the MRI may be a factor in the misclassifications of false-negative and false-positive.[19] Kim et al. (2024) assessed a deep learning triage application on a total of 947 MRI examinations performed in the emergency room and found sensitivity of 90%, specificity of 89%, accuracy of 89%, and an AUROC of 0.95. The present AI-assisted MRI had higher sensitivity and specificity, as well as a comparable AUROC of 0.933. Their research also showed that adding T2-weighted sequences was not helpful for the AI, highlighting the need for optimized DWI/FLAIR-based workflows for fast AIS detection.[20] In general, the present results are in line with current evidence on the high sensitivity and specificity of AI methods in the detection of acute ischemic lesions in MRI. At the same time, the variation between published studies, from substantially lower performance in some datasets to near-perfect results in others, indicates that algorithm performance is influenced by the characteristics of the training and validation populations, MRI protocols, lesion size and distribution, image quality, and reference-standard methodology. Therefore, these findings not only validate the use of AI-assisted MRI in diagnosis but also emphasize the importance of future multicenter and prospective studies for potential clinical adoption. LIMITATIONS There were some limitations to the study. The limited sample size and single center may make the results of this study difficult to generalize to other centers and other patient cohorts. The study also included a hypothetical data set, so the diagnostic estimates reported herein must be confirmed with prospectively collected clinical data. The performance of AI algorithms may be affected by changes in MRI protocols, MRI scanner settings, image quality, lesion size, and location of the infarcts. Furthermore, the interpretation of the radiology may be subject to differences of opinion among individual radiologists. Thus, external multicenter validation of larger patient cohorts and uniform MRI protocols is warranted.

AI-assisted MRI had a high diagnostic accuracy for acute ischemic stroke and offered a higher sensitivity, specificity, predictive values, and accuracy compared to the conventional radiological assessment. The significantly higher AUROC further indicated superior discriminatory ability. Overall, these results suggest that using AI alongside traditional radiological analysis could enhance the detection of acute ischemic stroke, especially in those that have mild imaging findings. Prospective multicenter studies are needed to validate these findings and assess their applicability in routine clinical practice.