A systematic review of the performance of Artificial Intelligence for automated DWI/FLAIR mismatch evaluation on MRI in ischemic stroke

A systematic review of the performance of Artificial Intelligence for automated DWI/FLAIR mismatch evaluation on MRI in ischemic stroke | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A systematic review of the performance of Artificial Intelligence for automated DWI/FLAIR mismatch evaluation on MRI in ischemic stroke Zahra Soltanali, Alireza Pourrahim, Chelsea Ruth-Ann Williams, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4494281/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The purpose of this study was to survey the existing artificial intelligence (AI) algorithms created for the automated detection of the diffusion-weighted imaging (DWI)–fluid-attenuated inversion recovery (FLAIR) mismatch and assess how their performance compares to that diagnostic techniques performed by neuroradiologists. The literature search for this systematic review was conducted in PubMed, MEDLINE, Ovid Embase, Web of Science, Scopus, and Cochrane databases up until February 2, 2024. The review team cross-checked the reference lists of the included studies to identify any additional relevant references, in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. We assessed the included studies using the Quality Assessment of Diagnostic Accuracy Studies 2 tool. The area under the curve (AUC) was reported in most studies, However, one study did not report this metric, The AI models achieved AUCs between 0.60 and 0.90, Sensitivity ranged from 0.6 to 0.9, and specificity ranged from 0.72 to 0.99, the positive predictive value (PPV), negative predictive value (NPV) and F1-Score were ranging from 0.72 to 0.93, 0.47 to 0.91 and 0.65 to 0.9, respectively. Additionally, the dice similarity coefficients (DSC) 0.73 & 0.8 were stated in two researches and accuracies ranging from 0.67 to 0.99. This review indicates that the current AI methods for DWI/FLAIR mismatch assessment may not be able to accurately determine the time since stroke onset based only on the DWI and FLAIR sequences, however, an AI-based approach focused on treatment eligibility, outcome prediction, and incorporating patient-specific information could potentially improve the care of stroke patients. artificial intelligence DWI/FLAIR mismatch stroke dice similarity coefficient MRI. Figures Figure 1 1. Introduction It is crucial to consider the proper sequence of MRI for multiple brain lesions, alongside the administration of intravenous contrast. This approach improves diagnostic accuracy and sensitivity compared to imaging without contrast ( 1 ). This method has the potential to aid in detecting lesions, determining their quantity, and identifying their precise locations ( 2 ). Rapid generation of weighted images using synthetic MR could aid in stroke imaging. These images are based on white matter T2/FLAIR hyper intensities (WMH) obtained from relaxometry data ( 3 ). Cavitation was independently predicted by DWI lesion size and the free water-corrected tissue mean diffusivity at baseline. Additionally, the within-lesion heterogeneity assessed using advanced diffusion imaging was indicative of tissue fate ( 4 ) ( 5 ). Utilizing FLAIR MRI datasets exclusively as a follow-up sequence led to reliable lesion segmentation results using the proposed method ( 6 ). However, the visibility of FLAIR hyper intensity lesions cannot definitively determine the time of stroke onset ( 7 ) ( 8 ). Given that one study aimed to examine the association between collateral status and DWI-FLAIR mismatch in acute ischemic stroke patients within the 4.5-hour time window; the results showed patients with early FLAIR lesion visibility exhibited better collateral circulation. The findings have important implications for managing stroke patients with unknown time-of-onset and support the idea of expanding the treatment window ( 9 ) ( 10 ). The patients who had an acute ischemic lesion on DWI but not on FLAIR imaging were presumably within a safe and effective thrombolysis time window ( 11 ). The employment of the MRI DWI/T2WI (T2-weighted images) mismatch to select AIS (Acute Ischemic Stroke) patients for intravenous tPA (tissue plasminogen activator) between 4.5 and 12 hours was proved to be both safe and effective. The results were comparable to using DWI/PWI or DWI/FLAIR mismatch for screening, yet the DWI/T2WI approach was faster and eliminated the necessity for the use of contrast agents ( 12 )( 13 ). Evaluating FLAIR-DWI mismatch and DWI volume in acute stroke patients could prove beneficial in predicting functional outcomes post-stroke ( 14 ). Several research studies have validated the capability of algorithms, in particular, DL (deep learning) algorithms, to identify the presence of ischemic lesions on both DWI and FLAIR sequences ( 15 ) ( 16 ). The use of transfer learning in conjunction with customized top CNN (convolutional neural network) models has proven to be impressive in its performance outcome for detecting both ischemic stroke and classifying its vascular territorial type based on DWI ( 17 ). The binary mismatch assessment, carried out by field experts such as neuroradiologists in clinical practice, has shown poor inter- and intra-observer agreement ( 7 ) ( 11 ) ( 18 ) ( 19 ) ( 20 ). The purpose of this Systematic Review was to explore the current AI algorithms designed for automated DWI/FLAIR mismatch assessment and assess their performance in comparison to the diagnostic accuracy of diagnostic techniques performed by neuroradiologists. 2. Methods The systematic review was carried out following the standards established in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement (21). 2.1. Literature search strategy and selection criteria Literatures were searched within the space of 3 months; December 2, 2023, to February 2, 2024, by using databases such as Embase, MEDLINE, Ovid Embase, Web of Science, Scopus, and Cochrane reference lists; which were cross-checked for additional references. Principal investigators of clinical trials with unpublished data were contacted for additional data. We searched articles published within the clinical and technical scope of the review. Search strings for each database can be found in Appendix A. The literatures were extracted, and the resulting references were submitted to a public reference manager (Mendeley 1.13.8, www.mendeley.com) to delete duplicate articles. Only studies that focused on machine learning methods that automatically identify DWI and FLAIR lesions on structural brain MRI radiomics features in AIS patients were included. The studies had to be peer-reviewed and written in English. Editorials, case series, letters, conference proceedings, reviews, papers that did not contain clinical data on humans and provided adequate details, papers that did not include experimental evidence, and inaccessible papers were excluded. Using selection criteria, we excluded 143 articles, and ultimately retained 11 articles for the final review (Fig 1). The studies had to meet a set of several criteria for inclusion, such as (1) Machine-learning algorithms designed to automatically recognize DWI and FLAIR lesions based on radiomic characteristics in structural brain MRI scans of ischemic stroke patients, (2) Automated DWI/FLAIR mismatch assessment using machine-learning techniques, and (3) Automated machine-learning techniques for determining the time since stroke onset based on DWI and FLAIR lesion mismatch. Investigations with criteria involving “Employment of machine learning models incorporating specialized MR sequences or modalities apart from DWI and FLAIR, or concentrating on either DWI or FLAIR exclusively”, “CTA-, CT- or US-based RA researches”, “Non-original studies (editorials, letters, reviews or opinions)”, “Insufficiency in detailing data acquisition, performance metrics or test strategy” were excluded. The two reviewers, medical students (Z.S. and A.P.), independently assessed all records using titles and abstracts, then proceeded to conduct a thorough examination of the potentially pertinent papers during the full-text review to determine final inclusion. A group consensus resolved all disagreements. 2.2. Data collection process The extraction of data from the included articles was performed by three reviewers (Z. S., A. P., and O.R.). The study characteristics extracted consisted of the following: The study location, first author, sample size, size of training and test data set, cohorts’ classification, population demographics, NIHSS (National Institutes of Health Stroke Scale) score during admission, the time interval between the onset of the symptoms and the MRI examination, and the percentage of cases where MRI was performed within 4.5 hours after the symptoms began. Additionally, the type of scanner, scanner characteristics, and the employed AI methods and its features (used slices, Stroke Lesion Volumes (mL), features, optimizer) were extracted. The performance data for the algorithms and the utilized comparator were extracted. They consisted of a validation method, Area under the curve (AUC), Accuracy (Acc), Sensitivity (sens), Specificity (spec), Negative predictive value (NPV), Precision or Positive predictive value (PPV), F1-score and Dice Similarity Coefficient (DSC). If there were multiple instances of performance data for different algorithms, only the metrics from the most effective algorithm were included unless otherwise indicated. 2.3. Quality Assessment of Included Studies Although there were limitations regarding study quantity and quality, the authors accepted these shortcomings without compromising the article's overall validity and relevance. The tool; the questionnaire Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) (22), was used independently by two authors (AP and ZS) to evaluate the diagnostic accuracy reported in the studies. The QUADAS-2 evaluated the concerns for applicability and risk of bias in the index test, the domains of the patients, the flow and timing, and the reference standard. Although it didn’t yield an overall numerical score, it provided a reproducible evaluation that categorized seven types of biases as “high”, “low”, or “unclear”. 3. Results 3.1. Study selection The literature search identified 3379 articles initially. Subsequent screening of 1757 titles and abstracts resulted in the selection of 143 articles considered relevant for full-text review, however, a few of them contained outcomes that were unrelated to this study; i.e., a focus on the independent detection of FLAIR and DWI lesions, without considering them as a mismatch, while exploring neurorehabilitation and unraveling the neural underpinnings of functional impairments (23)(24)(25) or the combination of some imaging methods. Some papers were centered on the exploration and enhancement of approaches for automatically identifying and localizing ischemic lesion zones or predicting the future state of ischemic tissue, with a focus on aiding doctors and researchers in their work (26)(27). As a result, a total of 12 articles were included in this review. The details of the inclusion process are outlined in Figure 1 . 3.2. Study characteristics and scanner parameters Every study was incorporated, with the exception of one, which was a prospective study (28). Image data from a range of hospitals or university databases within the designated timeframe were employed for the analysis. A summary of study characteristics and scanner specifications is presented in Table 1 . The study population ranged from 72 to 1098 patients who had been diagnosed with stroke. All five studies split their datasets into training and testing subsets, with the distribution of patients in the two groups ranging from 16% to 51%. In four studies, the test population had a significantly higher proportion, with distributions varying from 30% to 51%. (29)(30)(31)(32). However, in one research study, the distribution of patients among groups was not documented (28). Participants in the five studies had a median age between 58.6 and 70.8 years. The difference in NIHSS scores at admission between the training and test populations was relatively minor in the studies examined. Some articles showed a higher NIHSS score upon admission in the training population compared to the test population. (33)(34). The test population in some articles exhibited higher scores than the training population (30) (35) (36). Others reported an average score for the entire study population (29) (28) (37) (38). One research did not report NIHSS scores upon admission (31). Only the study by Moon et al. investigated the association between lesion metrics as independent variables and baseline NIHSS as the dependent variable in a single regression analysis. (28). Yu et al. examined the association between prognosis prediction as the independent variable and baseline NIHSS as the dependent variable based on the t/χ ratio compared between two groups defined by the Modified Rankin Scale (mRS) as Good prognosis (mRs ≤ 2) (n = 83) and Poor prognosis (mRS > 2) (n = 65) (32). For three of the studies, the mean time from stroke occurrence to MRI examination was high, more than 4.5 hours (32)(28)(29). The median time from stroke onset to MRI in the remaining studies fell within a relatively short range of 3.5 to 4.5 hours across both training and test groups, except for one study that did not present any details. Four of the studies reported the percentage of patients who underwent MRI within 4.5 hours of stroke onset (29)(33)(30)(31); across the training and test populations, the percentage varied from 64.55% to 20.31%. The comparator for the AI algorithm in five of the ten studies was the human assessment of mismatch (29) (33) (34) (35). Although the studies employed a range of AI models, one investigation uniquely compared the performance of an AI algorithm that was FLAIR-based against a human expert's evaluation (36). While some studies did not have a comparator, the scope of this review deemed their inclusion relevant. 3.3. Automated evaluation of Time since Stroke and Segmentation The studies by Lee et al., Jiang et al., and Zhu et al. utilized a DL algorithm in combination with standard ML (machine learning) algorithms within a two-step framework to assess the Text Similarity Score (TSS)(33)(30)(34). The defined segmentation methods are seen in Table 2. Lee et al., Jiang et al., and Yu et al studies identified Random Forest and a Support Vector machine with a radial kernel (svmRadial) as the top-performing machine learning models, respectively (32)(30)(35). Zhu et al. and Akay et al. adopted an ensemble methodology that incorporated outputs of the five most effective machine learning algorithms, including Random Forest, Support Vector Machine, and Light GBM (31)(29). Polson et al., Akay et al., and Zhang et al. developed a CNN that could both segment lesions on unspecified image data and classify the time since stroke within a single unified model. (31)(34)(33). In the Benzakoun et al study, the researchers used a trained deep learning model that originated from the Edge-aware Generative Adversarial Network (Ea-GAN) to create synthetic FLAIR images in the test set; it utilized source DWI and apparent diffusion coefficient (ADC) maps instead of real FLAIR images, to identify AIS within 4.5 hours (36). Lu et al found that combining lesion features with whole-brain features led to a more accurate model for predicting AIS onset time and they also identified the most effective model within each group of whole-brain features (38). The study assessed how consistent radiologists were in interpreting DWI-FLAIR mismatches, finding moderate agreement between radiologists with scores ranging from 0.43 to 0.76 across five studies (35)(33)(31)(34)(36). 3.4. Optimization process The studies by Polson et al. and Yu et al. employed a 5-fold cross-validation technique to tune the hyper parameters of their models during training (33) (32). There were also 10-fold cross-validations for two studies (38) (30), 4-fold cross-validation for one study (31), and a 6.5-fold cross-validation in one study (28). The validation methods used in some studies were not reported. In order to train the segmentation models, the authors minimized a novel loss function by using the Adam optimization algorithm (29) (33) (28), AdaBound optimizer (34), Medical Image Processing, Analysis and Visualization (MIPAV) (30) or the Stochastic Gradient Descent (SGD) (31). Jiang et al., Zhu et al. The segmentation models were trained by minimizing a loss function tuned with hyper parameters (29) (30). Akay et al. employed a convolutional auto encoder (CAE) to examine the effects of reconstruction on supervised learning for time-to-MRI prediction, during model training, they used a Stochastic Gradient Descent (SGD) optimizer with Nesterov momentum to optimize the model (31). 3.5. Study Performance During this evaluation, performance measures for the AI models were recorded. Additionally, any missing data points were inputted using existing data whenever possible. Table 2 shows the performance results for time since stroke classification, which are considered the most reliable metrics due to their use of out-of-distribution external validation datasets. AUC was reported in most studies; however, one study did not report this metric (28). The AI models achieved AUCs between 0.60 and 0.90, and the comparators in two studies had AUCs of 0.69 and 0.88(31)(36). Despite being dichotomous (yes/no) in six assessments [1], all of the studies included sensitivity and specificity. The sensitivity ranged from 0.6 to 0.9, and the specificity ranged from 0.72 to 0.99, which defined these terms as true positive (recall) and true negative rate respectively. The analysis of the top performing models revealed that in all the studies the specificity was slightly greater than the sensitivity. Details are provided in Table 2 For the studies, the positive predictive value (PPV), also known as the precision, negative predictive value (NPV), and F1-Score were reported or calculated with ranges of 0.72 to 0.93, 0.47 to 0.91 and 0.65 to 0.9 respectively. Additionally, DSC (dice similarity coefficient) was stated as 0.73 and 0.8 in two researches (29) (31). In the same studies, the accuracy ranged from 0.67 to 0.99. PPV and NPV were likely calculated for the AI method and comparators to align with the other studies reporting these metrics. While performance measures for the comparators were reported in six studies, some studies did not include NPV or PPV (33)(28)(32). The sensitivity and specificity ranged from 0.43 to 0.85 and 0.59 to 0.92 respectively. The accuracy ranged from 0.61 to 0.87. The PPV and NPV ranged from 0.38 to 0.89 and 0.44 to 0.93, respectively (29)(33)(35)(31)(36) but the F1-score ranged from 0.63 to 0.88 in four studies (29)(33)(36)(35). The DSC, however, was not included in any of the studies. These results suggested that AI specificity, sensitivity, F1-score, and accuracy achieved a higher accuracy in diagnosing the time since stroke compared to radiologists. The study found the AI model comparable to radiologists in terms of PPV and NPV for classifying strokes within a value of time since stroke being ±4.5 h. One of the studies, however, compared the performance of the synthetic FLAIR and real FLAIR sequences where they concluded that the synthetic FLAIR could replace the real FLAIR sequences (36). In the Yu et al study they constructed the ML models for prognostic classification and region of interest (ROI) (32). 3.6. Agreement levels among radiologists This study compares the level of agreement between radiologists in assessing DWI-FLAIR mismatch. The Benzakoun et al study highlighted a substantial interobserver reproducibility for both real and synthetic FLAIR imaging with a nearly perfect concordance, a kappa coefficient of 0.88 after reaching a consensus was calculated. A weighted kappa coefficient of 0.76 for Fazekas scores which showed the white matter hyperintensities concordance between real FLAIR and synthetic FLAIR was also calculated (36). In contrast, Polson et al research, evaluating Fleiss' kappa values of 0.460 for the internal dataset and 0.575 for the external dataset, indicated moderate levels of agreement among three radiologists when assessing the DWI-FLAIR mismatch. The average rate of agreement between the DL classifications and radiologist assessments was 0.41 indicating a similar level of agreement (33). In Akay et al study, the study found that incorporating DL predictions improved the interrater agreement between junior doctors in diagnosing stroke onset time. For seniors 1 and junior 2 the interrater agreement remained unchanged with a Cohen’s kappa of 0.57 for the PRE-FLAIR (31). Zhang et al reported an inter-reader agreement (Fleiss’ kappa) of 0.46 among all three radiologists (34), additionally, an initial human interobserver agreement of 73.2 % (with a κ of 0.43) was reported in the Lee et al study (35). Four studies reported moderate agreement between researchers interpreting unenhanced MRI scans (35) (33) (31) (34). However, the findings differed from those of Benzakoun et al. (36). 3.7. Quality Assessment The comprehensive analysis of the quality assessment of the studies, carried out using the QUADAS-2 tool, is outlined in Table 3 The results of the quality assessment indicate a varying degree of bias across the studies. While some studies demonstrated a low risk of bias and high applicability across all domains (e.g., Akay et al., 2023), others exhibited a high or unclear risk of bias in certain areas (e.g., Lu et al., 2024; Moon et al., 2022; Jiang et al., 2022; Zhu et al., 2021; Zhang et al., 2020; Lee et al., 2020). The most common area of high or unclear risk was in the domain of patient selection. In terms of applicability, most studies were found to have a low concern. However, a few studies had high or unclear concerns in the domains of patient selection and reference standards (e.g., Yu et al., 2023; Jiang et al., 2022; Zhu et al., 2021). This underscored the importance of choosing a reference standard that was highly applicable to the review question and ensured that the patient selection criteria were relevant and clearly defined. Only two of the studies (e.g., Lu et al., 2024; Moon et al., 2022) had smaller cohort sizes. However, despite their smaller scale, these studies still provided sufficient data for our study's emphasis on automated segmentation of stroke lesions on MRI images. In conclusion, while the overall quality of the included studies is satisfactory, there is room for improvement, particularly in the domains of patient selection and reference standards. Due to the limited number of eligible studies, a meta-analysis was not carried out as part of this review. 4. Discussion 4.1. Automated evaluation of time since stroke Alteplase, a clot-busting drug, has proved to be beneficial in patients with acute ischemic stroke when selected based on a DWI-FLAIR mismatch according to a randomized controlled trial. The provided research suggested that analyzing the relative signal intensity (rSI) on FLAIR MRI images (FLAIR-rSI) was able to estimate the median stroke onset time to be around 6.1 hours after midnight (IQR 2.7 h) ( 39 ). If the recognized symptoms lasted less than 4.5 hours, studies have signaled that thrombolysis may be feasible in selected patients with wake-up stroke (WUS) or unknown onset stroke (UOS) who have a partial FLAIR signal positivity ( 40 ). The increased use of IV tPA ( 41 ); although the AI-assisted DWI/FLAIR mismatch analysis currently in use relies on FLAIR lesion visibility as an indirect indicator of time since stroke onset, classified patients as either within or exceeding the 4.5-hour thrombolysis time frame. Patients with unclear stroke onset time were often considered ineligible for thrombolysis, even if the time since the occurrence of a stroke exceeded 6 h. Current clinical guidelines recommend the use of Multimodal Magnetic Resonance Imaging (MRI) to assess various mismatches in acute ischemic stroke patients. The utilization of AI-powered methods that evaluate diverse MRI features may serve to augment the accuracy, advance patient outcomes, and positively influence the clinical decision-making process in the management of AIS. However, maintaining a collaborative environment between clinical practitioners and academic investigators is crucial to actualizing the full potential of AI in optimizing the management of AIS ( 42 ). The AI algorithms utilized in the included studies performed equivalent to or even better than neuroradiologists and neurologists in the binary classification of time since stroke onset within or beyond 4.5 hours based on DWI/FLAIR imaging. 4.2. Artificial Intelligence vs. Human Readings Radiomic imaging features describing the infarct lesion on follow-up MRI sequences, such as DWI and T2-FLAIR, have been shown to be predictive of poor long-term functional outcomes in AIS patients. In addition, the combination of imaging-derived features and standard clinical data available up to 24 hours after symptom onset has been found to yield the most accurate predictive models for determining long-term functional outcomes in acute ischemic stroke ( 43 ). Automating stroke assessment by identifying slices, hemispheres, segmenting stroke regions in DWI, and measuring hypo-perfused tissue in Perfusion-weighted imaging (PWI) significantly improves consistency and accuracy in stroke evaluation ( 44 ). Combining multiple data modalities, including DWI, ADC maps, and clinical patient information, has allowed for the construction of AI algorithms that can accurately classify acute ischemic stroke subtypes ( 45 ). One study investigated the ability of the RAPID AI (RAPID) software to estimate the infarct size on DWI in vertebrobasilar ischemia (VBI) compared to supratentorial ischemia (STI). The analysis revealed that the RAPID AI platform was less effective at identifying and measuring infarct volumes of 2 ml or less, in contrast to the visual DWI assessment conducted by the neurologists ( 46 ). Researchers explored methods to enhance FLAIR image segmentation, including the application of an algorithm for intensity correction, the utilization of an automated filter to address inhomogeneity, and the implementation of approaches to minimize the adverse impact of high-intensity areas. The FLAIR lesion segmentation method demonstrated promising results in estimating lesion volumes when evaluated on a follow-up FLAIR imaging dataset ( 6 ). The AI tool was found to have superior sensitivity and specificity in detecting acute ischemic lesions on MRI compared to the assessment of an experienced neuroradiologist. Ischemic lesion properties dictate a test's sensitivity to identify true strokes. In contrast, good image quality is crucial for specificity, meaning the ability to rule out non-strokes ( 47 ). 4.3. Agreement assessment The observed variability in the interpretation of DWI-ASPECTS and DWI-FLAIR mismatch suggested that the agreement between them may not be adequate to ensure repeatable and consistent clinical decisions about mechanical thrombectomy ( 48 ). The study found that the concordance between the quantitative and qualitative methods used to quantify FLAIR and DWI signals was fair; (73%, κ = 0.44 for relative FLAIR (Rflair) and 74%, κ = 0.48 for relative DWI (rDWI) ( 49 )( 50 ). Although the majority of researchers reported that the DWI-FLAIR mismatch was the most difficult imaging characteristic to interpret, the agreement for this criterion was 74%, corresponding to a kappa of 0.60( 51 ). In addition, the study reported a moderate level of inter-rater agreement, with a kappa of 0.58, for the visual grading of the imaging parameters ( 52 ). In some studies, the inter- and intra-observer reliability fell within a substantial to excellent range (kappa values ranging from 0.63 to 1.00)( 53 ), or a 2.5-minute inter- and intra-rater agreement of κ = .88 − 1( 54 ). The analysis showed an almost moderate agreement across the protocols used in the studies, barring one study that demonstrated a particularly high level of agreement between an artificial intelligence-enhanced ultrafast (UF) protocol and the reference protocols in regards to the distribution, detection, number of AIS in DWI (κ = 0.98, 0.98, and 0.87 respectively), the presence of a parenchymal hyperintensity and presence of vascular hyper intensities in the AIS region in FLAIR (κ = 0.93 and 0.89 respectively) ( 55 ). 5. Limitation The approach provided two advantages in these situations: Firstly, the AI-powered model has demonstrated the ability to provide ratings with accuracy comparable to that of a human expert rater. Secondly, the AI system can augment the rating in cases where the assessment would otherwise be indeterminable, thereby improving the overall sensitivity and performance. The search results focused on the use of perfusion imaging in the evaluation and management of acute ischemic stroke ( 56 ). The search results showed the limitations imposed by the narrow focus of the inclusion criteria that were used. Another limitation of this review was the potential risk of publication bias, as it may not have been possible to rule out the possibility that studies with poor performance were not published; it would lead to a false and overly optimistic view of the potential and the performance of AI algorithms for a DWI/FLAIR mismatch assessment. The six reviewed methods focused on the classification of DWI/FLAIR mismatch determining time since stroke for AIS patients. In eight of the studies we reviewed, there was a clear specification of the stroke patient selection through well-defined inclusion and exclusion criteria. Notably, the presence of indiscernible lesions was a more included factor in the exclusion criteria (exception of ( 28 ) ( 29 )). Additionally, there was a lack of detailed demographic characteristics of the cohort that was noted ( 28 ). This lack of detail in the cohort characteristics could raise the risk of an unbalanced training and test cohort selection and impact the performance and generalizability of the models. The model may perform well on the test data but fails to generalize to new, unseen data. The reporting of the percentage of patients who had a stroke within 4.5 hours is inconsistent across the studies reviewed. This inconsistency could potentially introduce bias in the test data, leading to an overestimation of model performances and limiting the integration of these models into clinical practices. While the search was aimed to identify studies on DWI/FLAIR mismatch assessment, the review was not limited to excluding articles that also addressed outcome prediction and AI in prognosis performance. Ultimately, only a single study was found that met the defined criteria ( 32 ). 6. Conclusion The machine learning and deep learning algorithms investigated in this review demonstrated the ability to classify time since stroke within ± 4.5 hours with performance that was on par with or even surpassed that of domain experts. Given that the current AI methods for DWI/FLAIR mismatch assessment are most likely incapable of accurately determining the time since stroke onset by using only the DWI and FLAIR imaging sequences, focused on outcome prediction, treatment eligibility, and incorporating patient-specific data, could improve the management of stroke patients. Declarations Acknowledgments The authors wish to acknowledge all researchers that their publications were used in the present study and Ilam University of Medical Sciences is also thanked for its financial support. Ethical Approval Not applicable. Funding This work was supported by the Ilam University of Medical Sciences, Ilam, Iran (grant number: 9811117036). Availability of data and materials The data supporting the findings of this study are available from the corresponding author upon reasonable request from any investigator. Competing interests The authors declare that they have no conflict of interest. Author Contributions Conceptualization, Z.S. and O.R. Methodology, Z.S., I.A., and A.P.; Investigation, Z.S. and MH.E; Formal analysis, Z.S., and MH.E. Writing—original draft preparation, Z.S. and CH.W. Writing—review and editing Z.S., A.P. and CH..; Project administration, Z.S. and O.R. All authors have read and agreed to the published version of the manuscript. References Roozpeykar S, Azizian M, Zamani Z, Farzan MR, Veshnavei HA, Tavoosi N, et al. Contrast-enhanced weighted-T1 and FLAIR sequences in MRI of meningeal lesions. Am J Nucl Med Mol Imaging. 2022;12(2):63–70. Absinta M, Sati P, Reich DS. Advanced MRI and staging of multiple sclerosis lesions. Nat Rev Neurol. 2016 Jun;12(6):358–68. André J, Barrit S, Jissendi P. Synthetic MRI for stroke: a qualitative and quantitative pilot study. Sci Rep. 2022 Jul;12(1):11552. Duering M, Adam R, Wollenweber FA, Bayer-Karpinska A, Baykara E, Cubillos-Pinilla LY, et al. Within-lesion heterogeneity of subcortical DWI lesion evolution, and stroke outcome: A voxel-based analysis. J Cereb blood flow Metab Off J Int Soc Cereb Blood Flow Metab. 2020 Jul;40(7):1482–91. H Buck B, Akhtar N, Alrohimi A, Khan K, Shuaib A. Stroke mimics: incidence, aetiology, clinical features and treatment. Ann Med. 2021 Dec;53(1):420–36. Subbanna NK, Rajashekar D, Cheng B, Thomalla G, Fiehler J, Arbel T, et al. Stroke lesion segmentation in FLAIR MRI datasets using customized Markov random fields. Front Neurol [Internet]. 2019;10(MAY). Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85068002336&doi=10.3389%2Ffneur.2019.00541&partnerID=40&md5=c819ed2f1b665e0ddc24590b9c34f350 Emeriau S, Serre I, Toubas O, Pombourcq F, Oppenheim C, Pierot L. Can diffusion-weighted imaging-fluid-attenuated inversion recovery mismatch (positive diffusion-weighted imaging/negative fluid-attenuated inversion recovery) at 3 Tesla identify patients with stroke at <4.5 hours? Stroke. 2013 Jun;44(6):1647–51. Ebinger M, Galinovic I, Rozanski M, Brunecker P, Endres M, Fiebach JB. Fluid-attenuated inversion recovery evolution within 12 hours from stroke onset: a reliable tissue clock? Stroke. 2010 Feb;41(2):250–5. Berthezène Y, Eker O, Makris N, Bettan M, Mansuy A, Chabrol A, et al. Collateral circulation assessment within the 4.5 h time window in patients with and without DWI/FLAIR MRI mismatch. J Neurol Sci. 2018 Nov;394:94–8. Rimmele DL, Thomalla G. Wake-up stroke: clinical characteristics, imaging findings, and treatment option - an update. Front Neurol. 2014;5:35. Thomalla G, Cheng B, Ebinger M, Hao Q, Tourdias T, Wu O, et al. DWI-FLAIR mismatch for the identification of patients with acute ischaemic stroke within 4·5 h of symptom onset (PRE-FLAIR): a multicentre observational study. Lancet Neurol. 2011 Nov;10(11):978–86. Bai Q-K, Zhao Z-G, Lu L-J, Shen J, Zhang J-Y, Sui H-J, et al. Treating ischaemic stroke with intravenous tPA beyond 4.5 hours under the guidance of a MRI DWI/T2WI mismatch was safe and effective. Stroke Vasc Neurol. 2019 Mar;4(1):8–13. Thomalla G, Simonsen CZ, Boutitie F, Andersen G, Berthezene Y, Cheng B, et al. MRI-Guided Thrombolysis for Stroke with Unknown Time of Onset. N Engl J Med. 2018 Aug;379(7):611–22. Jiang L, Peng M, Geng W, Chen H, Su H, Zhao B, et al. FLAIR hyperintensities-DWI mismatch in acute stroke: associations with DWI volume and functional outcome. Brain Imaging Behav. 2020 Aug;14(4):1230–7. Wolman DN, van Ommen F, Tong E, Kauw F, Dankbaar JW, Bennink E, et al. Non-contrast dual-energy CT virtual ischemia maps accurately estimate ischemic core size in large-vessel occlusive stroke. Sci Rep. 2021 Mar;11(1):6745. Liu L, Kurgan L, Wu F-X, Wang J. Attention convolutional neural network for accurate segmentation and quantification of lesions in ischemic stroke disease. Med Image Anal. 2020 Oct;65:101791. Cetinoglu YK, Koska IO, Uluc ME, Gelal MF. Detection and vascular territorial classification of stroke on diffusion-weighted MRI by deep learning. Eur J Radiol. 2021 Dec;145:110050. Brinjikji W, Abbasi M, Arnold C, Benson JC, Braksick SA, Campeau N, et al. e-ASPECTS software improves interobserver agreement and accuracy of interpretation of aspects score. Interv Neuroradiol [Internet]. 2021;27(6):781–7. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85104440160&doi=10.1177%2F15910199211011861&partnerID=40&md5=85700d836ed76ec5a818c6bc62dde819 Weyland CS, Papanagiotou P, Schmitt N, Joly O, Bellot P, Mokli Y, et al. Hyperdense Artery Sign in Patients With Acute Ischemic Stroke-Automated Detection With Artificial Intelligence-Driven Software. Front Neurol. 2022;13:807145. Rahman S, Sarker S, Haque AKMN, Uttsha MM, Islam MF, Deb S. AI-Driven Stroke Rehabilitation Systems and Assessment: A Systematic Review. IEEE Trans neural Syst Rehabil Eng a Publ IEEE Eng Med Biol Soc. 2023;31:192–207. McInnes MDF, Moher D, Thombs BD, McGrath TA, Bossuyt PM, Clifford T, et al. Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies: The PRISMA-DTA Statement. JAMA. 2018 Jan;319(4):388–96. Whiting PF, Rutjes AWS, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011 Oct;155(8):529–36. Bian R, Huo M, Liu W, Mansouri N, Tanglay O, Young I, et al. Connectomics underlying motor functional outcomes in the acute period following stroke. Front Aging Neurosci. 2023;15:1131415. Bonkhoff AK, Schirmer MD, Bretzner M, Etherton M, Donahue K, Tuozzo C, et al. Abnormal dynamic functional connectivity is linked to recovery after acute ischemic stroke. Hum Brain Mapp [Internet]. 2021;42(7):2278–91. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85101879623&doi=10.1002%2Fhbm.25366&partnerID=40&md5=d925fcf7fbbb422add80eec0ab78c5f9 Gandolfi M, Galazzo IB, Pavan RG, Cruciani F, Vale N, Picelli A, et al. eXplainable AI Allows Predicting Upper Limb Rehabilitation Outcomes in Sub-Acute Stroke Patients. IEEE J Biomed Heal Informatics [Internet]. 2023;27(1):263–73. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85141606409&doi=10.1109%2FJBHI.2022.3220179&partnerID=40&md5=47174bb0378ea3178e20d18701ad7f96 Chang H-H, Yeh S-J, Chiang M-C, Hsieh S-T. RU-Net: skull stripping in rat brain MR images after ischemic stroke with rat U-Net. BMC Med Imaging. 2023 Mar;23(1):44. Huang S, Shen Q, Duong TQ. Artificial neural network prediction of ischemic tissue fate in acute stroke imaging. J Cereb Blood Flow Metab. 2010 Sep;30(9):1661–70. Moon HS, Heffron L, Mahzarnia A, Obeng-Gyasi B, Holbrook M, Badea CT, et al. Automated multimodal segmentation of acute ischemic stroke lesions on clinical MR images. Magn Reson Imaging. 2022 Oct;92:45–57. Zhu HC, Jiang L, Zhang H, Luo LM, Chen Y, Chen YC. An automatic machine learning approach for ischemic stroke onset time identification based on DWI and FLAIR imaging. NEUROIMAGE-CLINICAL. 2021;31. Jiang L, Wang S, Ai Z, Shen T, Zhang H, Duan S, et al. Development and external validation of a stability machine learning model to identify wake-up stroke onset time from MRI. Eur Radiol. 2022 Jun;32(6):3661–9. Akay EMZ, Rieger J, Schöttler R, Behland J, Schymczyk R, Khalil AA, et al. A deep learning analysis of stroke onset time prediction and comparison to DWI-FLAIR mismatch. NeuroImage Clin. 2023;40:103544. Yu H, Wang Z, Sun Y, Bo W, Duan K, Song C, et al. Prognosis of ischemic stroke predicted by machine learning based on multi-modal MRI radiomics. Front psychiatry. 2022;13:1105496. Polson JS, Zhang H, Nael K, Salamon N, Yoo BY, El-Saden S, et al. Identifying acute ischemic stroke patients within the thrombolytic treatment window using deep learning. J Neuroimaging. 2022 Nov;32(6):1153–60. Zhang H, Polson JS, Nael K, Salamon N, Yoo B, El-Saden S, et al. Intra-domain task-adaptive transfer learning to determine acute ischemic stroke onset time. Comput Med Imaging Graph. 2021 Jun;90:101926. Lee H, Lee E-J, Ham S, Lee H-B, Lee JS, Kwon SU, et al. Machine Learning Approach to Identify Stroke Within 4.5 Hours. Stroke. 2020 Mar;51(3):860–6. Benzakoun J, Deslys M-A, Legrand L, Hmeydia G, Turc G, Hassen W Ben, et al. Synthetic FLAIR as a Substitute for FLAIR Sequence in Acute Ischemic Stroke. Radiology. 2022 Apr;303(1):153–9. Zhang R, Wang J. Machine Learning-Based Prediction of Subsequent Vascular Events After 6 Months in Chinese Patients with Minor Ischemic Stroke. Int J Gen Med. 2022;15:3797–808. Lu J, Guo Y, Wang M, Luo Y, Zeng X, Miao X, et al. Determining acute ischemic stroke onset time using machine learning and radiomics features of infarct lesions and whole brain. Math Biosci Eng. 2024 Jan;21(1):34–48. Cheng B, Pinnschmidt H, Königsberg A, Schlemm E, Boutitie F, Ebinger M, et al. Estimating nocturnal stroke onset times by magnetic resonance imaging in the WAKE-UP trial. Int J stroke Off J Int Stroke Soc. 2022 Mar;17(3):323–30. Jakubicek S, Krebs S, Posekany A, Ferrari J, Szabo J, Siarnik P, et al. Modified DWI-FLAIR mismatch guided thrombolysis in unknown onset stroke. J Thromb Thrombolysis. 2019 Feb;47(2):167–73. Adil MM, Luby M, Lynch JK, Hsia AW, Kalaria CP, Nadareishvili Z, et al. Routine use of FLAIR-negative MRI in the treatment of unknown onset stroke. J stroke Cerebrovasc Dis Off J Natl Stroke Assoc. 2020 Sep;29(9):105093. Ben Alaya I, Limam H, Kraiem T. Automatic triaging of acute ischemic stroke patients for reperfusion therapies using Artificial Intelligence methods and multiple MRI features: A review. Clin Imaging. 2023 Dec;104:109992. Gerbasi A, Konduri P, Tolhuisen M, Cavalcante F, Rinkel L, Kappelhof M, et al. Prognostic Value of Combined Radiomic Features from Follow-Up DWI and T2-FLAIR in Acute Ischemic Stroke. J Cardiovasc Dev Dis. 2022 Dec;9(12). Ben Alaya I, Limam H, Kraiem T. Applications of artificial intelligence for DWI and PWI data processing in acute ischemic stroke: Current practices and future directions. Clin Imaging. 2022 Jan;81:79–86. Miyamoto N, Ueno Y, Yamashiro K, Hira K, Kijima C, Kitora N, et al. Stroke classification and treatment support system artificial intelligence for usefulness of stroke diagnosis. Front Neurol. 2023;14:1295642. Lakatos L, Bolognese M, Müller M, Österreich M, von Hessling A. Automated Supra- and Infratentorial Brain Infarct Volume Estimation on Diffusion Weighted Imaging Using the RAPID Software. Front Neurol. 2022;13:907151. Krag CH, Müller FC, Gandrup KL, Raaschou H, Andersen MB, Brejnebøl MW, et al. Diagnostic test accuracy study of a commercially available deep learning algorithm for ischemic lesion detection on brain MRIs in suspected stroke patients from a non-comprehensive stroke center. Eur J Radiol. 2023 Nov;168:111126. Fahed R, Lecler A, Sabben C, Khoury N, Ducroux C, Chalumeau V, et al. DWI-ASPECTS (Diffusion-Weighted Imaging-Alberta Stroke Program Early Computed Tomography Scores) and DWI-FLAIR (Diffusion-Weighted Imaging-Fluid Attenuated Inversion Recovery) Mismatch in Thrombectomy Candidates: An Intrarater and Interrater Agreement St. Stroke. 2018 Jan;49(1):223–7. Scheldeman L, Wouters A, Dupont P, Christensen S, Boutitie F, Cheng B, et al. Diffusion-Weighted Imaging and Fluid-Attenuated Inversion Recovery Quantification to Predict Diffusion-Weighted Imaging-Fluid-Attenuated Inversion Recovery Mismatch Status in Ischemic Stroke With Unknown Onset. Stroke. 2022 May;53(5):1665–73. Civrny J, Sedlackova Z, Malenak T, Kucera P, Machal D, Kocher M, et al. Comparison of semi-quantitative and visual assessment of early MRI signal evolution in acute ischaemic stroke. Eur J Radiol open. 2023;10:100488. Galinovic I, Dicken V, Heitz J, Klein J, Puig J, Guibernau J, et al. Homogeneous application of imaging criteria in a multicenter trial supported by investigator training: A report from the WAKE-UP study. Eur J Radiol. 2018 Jul;104:115–9. Regenhardt RW, Bretzner M, Zanon Zotin MC, Bonkhoff AK, Etherton MR, Hong S, et al. Radiomic signature of DWI-FLAIR mismatch in large vessel occlusion stroke. J neuroimaging Off J Am Soc Neuroimaging. 2022 Jan;32(1):63–7. Girot M, Leclerc X, Gauvrit J-Y, Verdelho A, Pruvo J-P, Leys D. Cerebral magnetic resonance imaging within 6 hours of stroke onset: inter- and intra-observer reproducibility. Cerebrovasc Dis. 2003;16(2):122–7. Kits A, Al-Saadi J, De Luca F, Janzon F, Mazya M V, Lundberg J, et al. 2.5-Minute Fast Brain MRI with Multiple Contrasts in Acute Ischemic Stroke. Neuroradiology. 2024 Mar; Verclytte S, Gnanih R, Verdun S, Feiweier T, Clifford B, Ambarki K, et al. Ultrafast MRI using deep learning echoplanar imaging for a comprehensive assessment of acute ischemic stroke. Eur Radiol. 2023 May;33(5):3715–25. Ho KC, Speier W, Zhang H, Scalzo F, El-Saden S, Arnold CW. A Machine Learning Approach for Classifying Ischemic Stroke Onset Time From Imaging. IEEE Trans Med Imaging. 2019 Jul;38(7):1666–76. Tables Tables 1 to 3 are available in the Supplementary Files section Additional Declarations No competing interests reported. Supplementary Files AppendixA.docx Tablesfinal.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4494281","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":314990346,"identity":"3b4897c2-272e-42ee-9f9e-1464dd3dbcca","order_by":0,"name":"Zahra Soltanali","email":"","orcid":"","institution":"Ilam University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Zahra","middleName":"","lastName":"Soltanali","suffix":""},{"id":314990347,"identity":"065eba66-b2e4-46ee-9814-48fc3d1db1b2","order_by":1,"name":"Alireza Pourrahim","email":"","orcid":"","institution":"Ilam University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Alireza","middleName":"","lastName":"Pourrahim","suffix":""},{"id":314990348,"identity":"f8012b30-b55f-4d39-996b-f7fd0246b403","order_by":2,"name":"Chelsea Ruth-Ann Williams","email":"","orcid":"","institution":"University of Porto","correspondingAuthor":false,"prefix":"","firstName":"Chelsea","middleName":"Ruth-Ann","lastName":"Williams","suffix":""},{"id":314990349,"identity":"9b12476f-81ff-4233-be73-073811f97c7e","order_by":3,"name":"Mohammad Hossain Ekvan","email":"","orcid":"","institution":"Ilam University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"Hossain","lastName":"Ekvan","suffix":""},{"id":314990350,"identity":"03434685-3799-4bc0-a9b7-534528fdac71","order_by":4,"name":"Iraj Ahmadi","email":"","orcid":"","institution":"Ilam University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Iraj","middleName":"","lastName":"Ahmadi","suffix":""},{"id":314990351,"identity":"84a9f20e-5c8f-4e01-8c29-11b932e78d4d","order_by":5,"name":"Omid Raiesi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIiWNgGAWjYLCCBAYGHn725gNApoQM8Voke44lgLTwEG+TwYwcAxBNWAs/A+/DDw/+1MkYSOR8fnWjxoKHgf3w0Q34tEg2sBtLJLYd5jHnebvNOucY0GE8aWk38LrnABuDRGLDAR7L9txtxjlsQC0SPGaEtDD/SPhTx2NwIOeZcc4/4rSwSSSwMfMYnMhhfpzbRoQWyWY2NguQX4CBbMac2yfBw0bIL/zsbcw3f/ypswdG5ePPOd/q5PjZDx/Dq4WBGcFkkwCTeJWj6/5AiupRMApGwSgYOQAAEbBBT8uPkKEAAAAASUVORK5CYII=","orcid":"","institution":"Ilam University of Medical Sciences","correspondingAuthor":true,"prefix":"","firstName":"Omid","middleName":"","lastName":"Raiesi","suffix":""}],"badges":[],"createdAt":"2024-05-29 04:48:50","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4494281/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4494281/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":58672831,"identity":"8ff33738-24bc-4201-848f-b2e4ba63160c","added_by":"auto","created_at":"2024-06-19 15:12:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":37901,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart PRISMA. The study inclusion process\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4494281/v1/79e46dcddc50dd6025fcebdc.png"},{"id":58672836,"identity":"f45967f0-5cc6-4a6e-ad75-1d8a6e66a1b2","added_by":"auto","created_at":"2024-06-19 15:12:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":506408,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4494281/v1/b2b22c1e-9d12-4524-95cf-632a7a63fbd6.pdf"},{"id":58672833,"identity":"67093f1e-538d-4f24-8ed2-1182b2ff5187","added_by":"auto","created_at":"2024-06-19 15:12:15","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":16746,"visible":true,"origin":"","legend":"","description":"","filename":"AppendixA.docx","url":"https://assets-eu.researchsquare.com/files/rs-4494281/v1/dd8c3c8735248dce2b8a3250.docx"},{"id":58672832,"identity":"6d62aa5d-13bf-480f-912a-099a1c0fdc31","added_by":"auto","created_at":"2024-06-19 15:12:15","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":33424,"visible":true,"origin":"","legend":"","description":"","filename":"Tablesfinal.docx","url":"https://assets-eu.researchsquare.com/files/rs-4494281/v1/bb894e8a4b3956700cba6cb0.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A systematic review of the performance of Artificial Intelligence for automated DWI/FLAIR mismatch evaluation on MRI in ischemic stroke","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIt is crucial to consider the proper sequence of MRI for multiple brain lesions, alongside the administration of intravenous contrast. This approach improves diagnostic accuracy and sensitivity compared to imaging without contrast (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). This method has the potential to aid in detecting lesions, determining their quantity, and identifying their precise locations (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Rapid generation of weighted images using synthetic MR could aid in stroke imaging. These images are based on white matter T2/FLAIR hyper intensities (WMH) obtained from relaxometry data (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Cavitation was independently predicted by DWI lesion size and the free water-corrected tissue mean diffusivity at baseline. Additionally, the within-lesion heterogeneity assessed using advanced diffusion imaging was indicative of tissue fate (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Utilizing FLAIR MRI datasets exclusively as a follow-up sequence led to reliable lesion segmentation results using the proposed method (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). However, the visibility of FLAIR hyper intensity lesions cannot definitively determine the time of stroke onset (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e) (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Given that one study aimed to examine the association between collateral status and DWI-FLAIR mismatch in acute ischemic stroke patients within the 4.5-hour time window; the results showed patients with early FLAIR lesion visibility exhibited better collateral circulation. The findings have important implications for managing stroke patients with unknown time-of-onset and support the idea of expanding the treatment window (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e) (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). The patients who had an acute ischemic lesion on DWI but not on FLAIR imaging were presumably within a safe and effective thrombolysis time window (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). The employment of the MRI DWI/T2WI (T2-weighted images) mismatch to select AIS (Acute Ischemic Stroke) patients for intravenous tPA (tissue plasminogen activator) between 4.5 and 12 hours was proved to be both safe and effective. The results were comparable to using DWI/PWI or DWI/FLAIR mismatch for screening, yet the DWI/T2WI approach was faster and eliminated the necessity for the use of contrast agents (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e)(\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Evaluating FLAIR-DWI mismatch and DWI volume in acute stroke patients could prove beneficial in predicting functional outcomes post-stroke (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). Several research studies have validated the capability of algorithms, in particular, DL (deep learning) algorithms, to identify the presence of ischemic lesions on both DWI and FLAIR sequences (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e) (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). The use of transfer learning in conjunction with customized top CNN (convolutional neural network) models has proven to be impressive in its performance outcome for detecting both ischemic stroke and classifying its vascular territorial type based on DWI (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe binary mismatch assessment, carried out by field experts such as neuroradiologists in clinical practice, has shown poor inter- and intra-observer agreement (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e) (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e) (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e) (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e) (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). The purpose of this Systematic Review was to explore the current AI algorithms designed for automated DWI/FLAIR mismatch assessment and assess their performance in comparison to the diagnostic accuracy of diagnostic techniques performed by neuroradiologists.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cp\u003eThe systematic review was carried out following the standards established in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement (21).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.1. Literature search strategy and selection criteria \u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLiteratures were searched within the space of 3 months; December 2, 2023, to February 2, 2024, by using databases such as Embase, MEDLINE, Ovid Embase, Web of Science, Scopus, and Cochrane reference lists; which were cross-checked for additional references. Principal investigators of clinical trials with unpublished data were contacted for additional data. We searched articles published within the clinical and technical scope of the review. Search strings for each database can be found in Appendix A.\u003c/p\u003e\n\u003cp\u003eThe literatures were extracted, and the resulting references were submitted to a public reference manager (Mendeley 1.13.8, www.mendeley.com) to delete duplicate articles. Only studies that focused on machine learning methods that automatically identify DWI and FLAIR lesions on structural brain MRI radiomics features in AIS patients were included. The studies had to be peer-reviewed and written in English. Editorials, case series, letters, conference proceedings, reviews, papers that did not contain clinical data on humans and provided adequate details, papers that did not include experimental evidence, and inaccessible papers were excluded. Using selection criteria, we excluded 143 articles, and ultimately retained 11 articles for the final review \u003cstrong\u003e(Fig 1).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe studies had to meet a set of several criteria for inclusion, such as (1) Machine-learning algorithms designed to automatically recognize DWI and FLAIR lesions based on radiomic characteristics in structural brain MRI scans of ischemic stroke patients, (2) Automated DWI/FLAIR mismatch assessment using machine-learning techniques, and (3) Automated machine-learning techniques for determining the time since stroke onset based on DWI and FLAIR lesion mismatch. Investigations with criteria involving “Employment of machine learning models incorporating specialized MR sequences or modalities apart from DWI and FLAIR, or concentrating on either DWI or FLAIR exclusively”, “CTA-, CT- or US-based RA researches”, “Non-original studies (editorials, letters, reviews or opinions)”, “Insufficiency in detailing data acquisition, performance metrics or test strategy” were excluded. The two reviewers, medical students (Z.S. and A.P.), independently assessed all records using titles and abstracts, then proceeded to conduct a thorough examination of the potentially pertinent papers during the full-text review to determine final inclusion. A group consensus resolved all disagreements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.2. Data collection process\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe extraction of data from the included articles was performed by three reviewers (Z. S., A. P., and O.R.). The study characteristics extracted consisted of the following: The study location, first author, sample size, size of training and test data set, cohorts’ classification, population demographics, NIHSS (National Institutes of Health Stroke Scale) score during admission, the time interval between the onset of the symptoms and the MRI examination, and the percentage of cases where MRI was performed within 4.5 hours after the symptoms began. Additionally, the type of scanner, scanner characteristics, and the employed AI methods and its features (used slices, Stroke Lesion Volumes (mL), features, optimizer) were extracted. The performance data for the algorithms and the utilized comparator were extracted. They consisted of a validation method, Area under the curve (AUC), Accuracy (Acc), Sensitivity (sens), Specificity (spec), Negative predictive value (NPV), Precision or Positive predictive value (PPV), F1-score and Dice Similarity Coefficient (DSC). If there were multiple instances of performance data for different algorithms, only the metrics from the most effective algorithm were included unless otherwise indicated.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e2.3. Quality Assessment of Included Studies\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAlthough there were limitations regarding study quantity and quality, the authors accepted these shortcomings without compromising the article's overall validity and relevance. The tool; the questionnaire Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) (22), was used independently by two authors (AP and ZS) to evaluate the diagnostic accuracy reported in the studies. The QUADAS-2 evaluated the concerns for applicability and risk of bias in the index test, the domains of the patients, the flow and timing, and the reference standard. Although it didn’t yield an overall numerical score, it provided a reproducible evaluation that categorized seven types of biases as “high”, “low”, or “unclear”.\u003c/p\u003e"},{"header":"3. Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.1. Study selection\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe literature search identified 3379 articles initially. Subsequent screening of 1757 titles and abstracts resulted in the selection of 143 articles considered relevant for full-text review, however, a few of them contained outcomes that were unrelated to this study; i.e., a focus on the independent detection of FLAIR and DWI lesions, without considering them as a mismatch, while exploring neurorehabilitation and unraveling the neural underpinnings of functional impairments (23)(24)(25) or the combination of some imaging methods. Some papers were centered on the exploration and enhancement of approaches for automatically identifying and localizing ischemic lesion zones or predicting the future state of ischemic tissue, with a focus on aiding doctors and researchers in their work (26)(27). As a result,\u0026nbsp;a\u0026nbsp;total of 12 articles were included in this review. The details of the inclusion process are outlined in \u003cstrong\u003eFigure 1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.2. Study characteristics and scanner parameters\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEvery study was incorporated, with the exception of one, which was a prospective study (28). Image data from a range of hospitals or university databases within the designated timeframe were employed for the analysis. A summary of study characteristics and scanner specifications is presented in \u003cstrong\u003eTable 1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003eThe study population ranged from 72 to 1098 patients who had been diagnosed with stroke. All five studies split their datasets into training and testing subsets, with the distribution of patients in the two groups ranging from 16% to 51%. In four studies, the test population had a significantly higher proportion, with distributions varying from 30% to 51%. (29)(30)(31)(32). However, in one research study, the distribution of patients among groups was not documented (28).\u0026nbsp;Participants in the five studies had a median age between 58.6 and 70.8 years. The difference in NIHSS scores at admission between the training and test populations was relatively minor in the studies examined. Some articles showed a higher NIHSS score upon admission in the training population compared to the test population. (33)(34). The test population in some articles exhibited higher scores than the training population (30) (35) (36). Others reported an average score for the entire study population (29) (28) (37) (38). One research did not report NIHSS scores upon admission (31). Only the study by Moon et al. investigated the association between lesion metrics as independent variables and baseline NIHSS as the dependent variable in a single regression analysis. (28). Yu et al. examined the association between prognosis prediction as the independent variable and baseline NIHSS as the dependent variable based on the t/χ ratio compared between two groups defined by the Modified Rankin Scale (mRS) as Good prognosis (mRs ≤ 2) (n = 83) and Poor prognosis (mRS \u0026gt; 2) (n = 65) (32).\u0026nbsp;For three of the studies, the mean time from stroke occurrence to MRI examination was high, more than 4.5 hours (32)(28)(29). The median time from stroke onset to MRI in the remaining studies fell within a relatively short range of 3.5 to 4.5 hours across both training and test groups, except for one study that did not present any details. Four of the studies reported the percentage of patients who underwent MRI within 4.5 hours of stroke onset (29)(33)(30)(31); across the training and test populations, the percentage varied from 64.55% to 20.31%. The comparator for the AI algorithm in five of the ten studies was the human assessment of mismatch (29) (33) (34) (35). Although the studies employed a range of AI models, one investigation uniquely compared the performance of an AI algorithm that was FLAIR-based against a human expert's evaluation (36). While some studies did not have a comparator, the scope of this review deemed their inclusion relevant.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3. Automated evaluation of Time since Stroke and Segmentation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe studies by Lee et al., Jiang et al., and Zhu et al. utilized a DL algorithm in combination with standard ML (machine learning) algorithms within a two-step framework to assess the Text Similarity Score (TSS)(33)(30)(34). The defined segmentation methods are seen in \u003cstrong\u003eTable 2.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLee et al., Jiang et al., and Yu et al studies identified Random Forest and a Support Vector machine with a radial kernel (svmRadial) as the top-performing machine learning models, respectively (32)(30)(35). Zhu et al. and Akay et al. adopted an ensemble methodology that incorporated outputs of the five most effective machine learning algorithms, including Random Forest, Support Vector Machine, and Light GBM (31)(29). Polson et al., Akay et al., and Zhang et al. developed a CNN that could\u0026nbsp;both segment lesions on unspecified image data and classify the time since stroke within a single unified model. (31)(34)(33). In the Benzakoun et al study, the researchers used a trained deep learning model that originated from the Edge-aware Generative Adversarial Network (Ea-GAN) to create synthetic FLAIR images in the test set; it utilized source DWI and apparent diffusion coefficient (ADC) maps instead of real FLAIR images, to identify AIS within 4.5 hours (36). \u0026nbsp;Lu et al found that combining lesion features with whole-brain features led to a more accurate model for predicting AIS onset time and they also identified the most effective model within each group of whole-brain features (38). The study assessed how consistent radiologists were in interpreting DWI-FLAIR mismatches, finding moderate agreement between radiologists with scores ranging from 0.43 to 0.76 across five studies (35)(33)(31)(34)(36).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.4. Optimization process\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe studies by Polson et al. and Yu et al. employed a 5-fold cross-validation technique to tune the hyper parameters of their models during training (33) (32). There were also 10-fold cross-validations for two studies (38) (30), 4-fold cross-validation for one study (31), and a 6.5-fold cross-validation in one study (28).\u0026nbsp;The validation methods used in some studies were not reported. In order to train the segmentation models, the authors minimized a novel loss function by using the Adam optimization algorithm (29) (33) (28), AdaBound optimizer (34), Medical Image Processing, Analysis and Visualization (MIPAV) (30) or the Stochastic Gradient Descent (SGD) (31). Jiang et al., Zhu et al. The segmentation models were trained by minimizing a loss function tuned with hyper parameters (29) (30). Akay et al. employed a convolutional auto encoder (CAE) to examine the effects of reconstruction on supervised learning for time-to-MRI prediction, during model training, they used a Stochastic Gradient Descent (SGD) optimizer with Nesterov momentum to optimize the model (31).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.5. Study Performance\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring this evaluation, performance measures for the AI models were recorded. Additionally, any missing data points were inputted using existing data whenever possible. Table 2 shows the performance results for time since stroke classification, which are considered the most reliable metrics due to their use of out-of-distribution external validation datasets. AUC was reported in most studies; however, one study did not report this metric (28). The AI models achieved AUCs between 0.60 and 0.90, and the comparators in two studies had AUCs of 0.69 and 0.88(31)(36). Despite being dichotomous (yes/no) in six assessments [1], all of the studies included sensitivity and specificity. The sensitivity ranged from 0.6 to 0.9, and the specificity ranged from 0.72 to 0.99, which defined these terms as true positive (recall) and true negative rate respectively. The analysis of the top performing models revealed that in all the studies the specificity was slightly greater than the sensitivity. Details are provided in \u003cstrong\u003eTable 2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the studies, the positive predictive value (PPV), also known as the precision, negative predictive value (NPV), and F1-Score were reported or calculated with ranges of 0.72 to 0.93, 0.47 to 0.91 and 0.65 to 0.9 respectively. Additionally, DSC (dice similarity coefficient) was stated as 0.73 and 0.8 in two researches (29) (31). In the same studies, the accuracy ranged from 0.67 to 0.99. PPV and NPV were likely calculated for the AI method and comparators to align with the other studies reporting these metrics. While performance measures for the comparators were reported in six studies, some studies did not include NPV or PPV (33)(28)(32). The sensitivity and specificity ranged from 0.43 to 0.85 and 0.59 to 0.92 respectively. The accuracy ranged from 0.61 to 0.87. The PPV and NPV ranged from 0.38 to 0.89 and 0.44 to 0.93, respectively (29)(33)(35)(31)(36) but the F1-score ranged from 0.63 to 0.88 in four studies (29)(33)(36)(35). The DSC, however, was not included in any of the studies. These results suggested that AI specificity, sensitivity, F1-score, and accuracy achieved a higher accuracy in diagnosing the time since stroke compared to radiologists. The study found the AI model comparable to radiologists in terms of PPV and NPV for classifying strokes within a value of time since stroke being ±4.5 h. One of the studies, however, compared the performance of the synthetic FLAIR and real FLAIR sequences where they concluded that the synthetic FLAIR could replace the real FLAIR sequences (36). In the Yu et al study they constructed the ML models for prognostic classification and region of interest (ROI) (32).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.6. Agreement levels among radiologists\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study compares the level of agreement between radiologists in assessing DWI-FLAIR mismatch. The Benzakoun et al study highlighted a substantial interobserver reproducibility for both real and synthetic FLAIR imaging with a nearly perfect concordance, a kappa coefficient of 0.88 after reaching a consensus was calculated. A weighted kappa coefficient of 0.76 for Fazekas scores which showed the white matter hyperintensities concordance between real FLAIR and synthetic FLAIR was also calculated (36). In contrast, Polson et al research, evaluating Fleiss' kappa values of 0.460 for the internal dataset and 0.575 for the external dataset, indicated moderate levels of agreement among three radiologists when assessing the DWI-FLAIR mismatch. The average rate of agreement between the DL classifications and radiologist assessments was 0.41 indicating a similar level of agreement (33). In Akay et al study, the study found that incorporating DL predictions improved the interrater agreement between junior doctors in diagnosing stroke onset time. For seniors 1 and junior 2 the interrater agreement remained unchanged with a Cohen’s kappa of 0.57 for the PRE-FLAIR (31). \u0026nbsp;Zhang et al reported an inter-reader agreement (Fleiss’ kappa) of 0.46 among all three radiologists (34), additionally, an initial human interobserver agreement of 73.2 % (with a κ of 0.43) was reported in the Lee et al study (35). Four studies reported moderate agreement between researchers interpreting unenhanced MRI scans (35) (33) (31) (34). However, the findings differed from those of Benzakoun et al. (36).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.7. Quality Assessment\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe comprehensive analysis of the quality assessment of the studies, carried out using the QUADAS-2 tool, is outlined in \u003cstrong\u003eTable 3\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe results of the quality assessment indicate a varying degree of bias across the studies. While some studies demonstrated a low risk of bias and high applicability across all domains (e.g., Akay et al., 2023), others exhibited a high or unclear risk of bias in certain areas (e.g., Lu et al., 2024; Moon et al., 2022; Jiang et al., 2022; Zhu et al., 2021; Zhang et al., 2020; Lee et al., 2020). The most common area of high or unclear risk was in the domain of patient selection. In terms of applicability, most studies were found to have a low concern. However, a few studies had high or unclear concerns in the domains of patient selection and reference standards (e.g., Yu et al., 2023; Jiang et al., 2022; Zhu et al., 2021). This underscored the importance of choosing a reference standard that was highly applicable to the review question and ensured that the patient selection criteria were relevant and clearly defined.\u003c/p\u003e\n\u003cp\u003eOnly two of the studies (e.g., Lu et al., 2024; Moon et al., 2022) had smaller cohort sizes. However, despite their smaller scale, these studies still provided sufficient data for our study's emphasis on automated segmentation of stroke lesions on MRI images. In conclusion, while the overall quality of the included studies is satisfactory, there is room for improvement, particularly in the domains of patient selection and reference standards. Due to the limited number of eligible studies, a meta-analysis was not carried out as part of this review.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Automated evaluation of time since stroke\u003c/h2\u003e \u003cp\u003eAlteplase, a clot-busting drug, has proved to be beneficial in patients with acute ischemic stroke when selected based on a DWI-FLAIR mismatch according to a randomized controlled trial. The provided research suggested that analyzing the relative signal intensity (rSI) on FLAIR MRI images (FLAIR-rSI) was able to estimate the median stroke onset time to be around 6.1 hours after midnight (IQR 2.7 h) (\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e). If the recognized symptoms lasted less than 4.5 hours, studies have signaled that thrombolysis may be feasible in selected patients with wake-up stroke (WUS) or unknown onset stroke (UOS) who have a partial FLAIR signal positivity (\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e). The increased use of IV tPA (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e); although the AI-assisted DWI/FLAIR mismatch analysis currently in use relies on FLAIR lesion visibility as an indirect indicator of time since stroke onset, classified patients as either within or exceeding the 4.5-hour thrombolysis time frame. Patients with unclear stroke onset time were often considered ineligible for thrombolysis, even if the time since the occurrence of a stroke exceeded 6 h. Current clinical guidelines recommend the use of Multimodal Magnetic Resonance Imaging (MRI) to assess various mismatches in acute ischemic stroke patients. The utilization of AI-powered methods that evaluate diverse MRI features may serve to augment the accuracy, advance patient outcomes, and positively influence the clinical decision-making process in the management of AIS. However, maintaining a collaborative environment between clinical practitioners and academic investigators is crucial to actualizing the full potential of AI in optimizing the management of AIS (\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e). The AI algorithms utilized in the included studies performed equivalent to or even better than neuroradiologists and neurologists in the binary classification of time since stroke onset within or beyond 4.5 hours based on DWI/FLAIR imaging.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Artificial Intelligence vs. Human Readings\u003c/h2\u003e \u003cp\u003eRadiomic imaging features describing the infarct lesion on follow-up MRI sequences, such as DWI and T2-FLAIR, have been shown to be predictive of poor long-term functional outcomes in AIS patients. In addition, the combination of imaging-derived features and standard clinical data available up to 24 hours after symptom onset has been found to yield the most accurate predictive models for determining long-term functional outcomes in acute ischemic stroke (\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e). Automating stroke assessment by identifying slices, hemispheres, segmenting stroke regions in DWI, and measuring hypo-perfused tissue in Perfusion-weighted imaging (PWI) significantly improves consistency and accuracy in stroke evaluation (\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e). Combining multiple data modalities, including DWI, ADC maps, and clinical patient information, has allowed for the construction of AI algorithms that can accurately classify acute ischemic stroke subtypes (\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e). One study investigated the ability of the RAPID AI (RAPID) software to estimate the infarct size on DWI in vertebrobasilar ischemia (VBI) compared to supratentorial ischemia (STI). The analysis revealed that the RAPID AI platform was less effective at identifying and measuring infarct volumes of 2 ml or less, in contrast to the visual DWI assessment conducted by the neurologists (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e). Researchers explored methods to enhance FLAIR image segmentation, including the application of an algorithm for intensity correction, the utilization of an automated filter to address inhomogeneity, and the implementation of approaches to minimize the adverse impact of high-intensity areas. The FLAIR lesion segmentation method demonstrated promising results in estimating lesion volumes when evaluated on a follow-up FLAIR imaging dataset (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). The AI tool was found to have superior sensitivity and specificity in detecting acute ischemic lesions on MRI compared to the assessment of an experienced neuroradiologist. Ischemic lesion properties dictate a test's sensitivity to identify true strokes. In contrast, good image quality is crucial for specificity, meaning the ability to rule out non-strokes (\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Agreement assessment\u003c/h2\u003e \u003cp\u003eThe observed variability in the interpretation of DWI-ASPECTS and DWI-FLAIR mismatch suggested that the agreement between them may not be adequate to ensure repeatable and consistent clinical decisions about mechanical thrombectomy (\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e). The study found that the concordance between the quantitative and qualitative methods used to quantify FLAIR and DWI signals was fair; (73%, κ\u0026thinsp;=\u0026thinsp;0.44 for relative FLAIR (Rflair) and 74%, κ\u0026thinsp;=\u0026thinsp;0.48 for relative DWI (rDWI) (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e)(\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e). Although the majority of researchers reported that the DWI-FLAIR mismatch was the most difficult imaging characteristic to interpret, the agreement for this criterion was 74%, corresponding to a kappa of 0.60(\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e). In addition, the study reported a moderate level of inter-rater agreement, with a kappa of 0.58, for the visual grading of the imaging parameters (\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e). In some studies, the inter- and intra-observer reliability fell within a substantial to excellent range (kappa values ranging from 0.63 to 1.00)(\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e), or a 2.5-minute inter- and intra-rater agreement of κ\u0026thinsp;=\u0026thinsp;.88\u0026thinsp;\u0026minus;\u0026thinsp;1(\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e). The analysis showed an almost moderate agreement across the protocols used in the studies, barring one study that demonstrated a particularly high level of agreement between an artificial intelligence-enhanced ultrafast (UF) protocol and the reference protocols in regards to the distribution, detection, number of AIS in DWI (κ\u0026thinsp;=\u0026thinsp;0.98, 0.98, and 0.87 respectively), the presence of a parenchymal hyperintensity and presence of vascular hyper intensities in the AIS region in FLAIR (κ\u0026thinsp;=\u0026thinsp;0.93 and 0.89 respectively) (\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Limitation","content":"\u003cp\u003eThe approach provided two advantages in these situations: Firstly, the AI-powered model has demonstrated the ability to provide ratings with accuracy comparable to that of a human expert rater. Secondly, the AI system can augment the rating in cases where the assessment would otherwise be indeterminable, thereby improving the overall sensitivity and performance. The search results focused on the use of perfusion imaging in the evaluation and management of acute ischemic stroke (\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e). The search results showed the limitations imposed by the narrow focus of the inclusion criteria that were used. Another limitation of this review was the potential risk of publication bias, as it may not have been possible to rule out the possibility that studies with poor performance were not published; it would lead to a false and overly optimistic view of the potential and the performance of AI algorithms for a DWI/FLAIR mismatch assessment. The six reviewed methods focused on the classification of DWI/FLAIR mismatch determining time since stroke for AIS patients. In eight of the studies we reviewed, there was a clear specification of the stroke patient selection through well-defined inclusion and exclusion criteria. Notably, the presence of indiscernible lesions was a more included factor in the exclusion criteria (exception of (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e) (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e)). Additionally, there was a lack of detailed demographic characteristics of the cohort that was noted (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e). This lack of detail in the cohort characteristics could raise the risk of an unbalanced training and test cohort selection and impact the performance and generalizability of the models. The model may perform well on the test data but fails to generalize to new, unseen data. The reporting of the percentage of patients who had a stroke within 4.5 hours is inconsistent across the studies reviewed. This inconsistency could potentially introduce bias in the test data, leading to an overestimation of model performances and limiting the integration of these models into clinical practices. While the search was aimed to identify studies on DWI/FLAIR mismatch assessment, the review was not limited to excluding articles that also addressed outcome prediction and AI in prognosis performance. Ultimately, only a single study was found that met the defined criteria (\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e).\u003c/p\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThe machine learning and deep learning algorithms investigated in this review demonstrated the ability to classify time since stroke within \u0026plusmn;\u0026thinsp;4.5 hours with performance that was on par with or even surpassed that of domain experts. Given that the current AI methods for DWI/FLAIR mismatch assessment are most likely incapable of accurately determining the time since stroke onset by using only the DWI and FLAIR imaging sequences, focused on outcome prediction, treatment eligibility, and incorporating patient-specific data, could improve the management of stroke patients.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors wish to acknowledge all researchers that their publications were used in the present study and Ilam University of Medical Sciences is also thanked for its financial support.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Ilam University of Medical Sciences, Ilam, Iran (grant number: 9811117036).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data supporting the findings of this study are available from the corresponding author upon reasonable request from any investigator.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAuthor Contributions\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization, Z.S. and O.R. Methodology, Z.S., I.A., and A.P.; Investigation, Z.S. and MH.E; Formal analysis, Z.S., and MH.E. Writing\u0026mdash;original draft preparation, Z.S. and CH.W. Writing\u0026mdash;review and editing Z.S., A.P. and CH..; Project administration, Z.S. and O.R. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eRoozpeykar S, Azizian M, Zamani Z, Farzan MR, Veshnavei HA, Tavoosi N, et al. Contrast-enhanced weighted-T1 and FLAIR sequences in MRI of meningeal lesions. Am J Nucl Med Mol Imaging. 2022;12(2):63\u0026ndash;70.\u003c/li\u003e\n \u003cli\u003eAbsinta M, Sati P, Reich DS. Advanced MRI and staging of multiple sclerosis lesions. Nat Rev Neurol. 2016 Jun;12(6):358\u0026ndash;68.\u003c/li\u003e\n \u003cli\u003eAndr\u0026eacute; J, Barrit S, Jissendi P. Synthetic MRI for stroke: a qualitative and quantitative pilot study. Sci Rep. 2022 Jul;12(1):11552.\u003c/li\u003e\n \u003cli\u003eDuering M, Adam R, Wollenweber FA, Bayer-Karpinska A, Baykara E, Cubillos-Pinilla LY, et al. Within-lesion heterogeneity of subcortical DWI lesion evolution, and stroke outcome: A voxel-based analysis. J Cereb blood flow Metab Off J Int Soc Cereb Blood Flow Metab. 2020 Jul;40(7):1482\u0026ndash;91.\u003c/li\u003e\n \u003cli\u003eH Buck B, Akhtar N, Alrohimi A, Khan K, Shuaib A. Stroke mimics: incidence, aetiology, clinical features and treatment. Ann Med. 2021 Dec;53(1):420\u0026ndash;36.\u003c/li\u003e\n \u003cli\u003eSubbanna NK, Rajashekar D, Cheng B, Thomalla G, Fiehler J, Arbel T, et al. Stroke lesion segmentation in FLAIR MRI datasets using customized Markov random fields. Front Neurol [Internet]. 2019;10(MAY). Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85068002336\u0026amp;doi=10.3389%2Ffneur.2019.00541\u0026amp;partnerID=40\u0026amp;md5=c819ed2f1b665e0ddc24590b9c34f350\u003c/li\u003e\n \u003cli\u003eEmeriau S, Serre I, Toubas O, Pombourcq F, Oppenheim C, Pierot L. Can diffusion-weighted imaging-fluid-attenuated inversion recovery mismatch (positive diffusion-weighted imaging/negative fluid-attenuated inversion recovery) at 3 Tesla identify patients with stroke at \u0026lt;4.5 hours? Stroke. 2013 Jun;44(6):1647\u0026ndash;51.\u003c/li\u003e\n \u003cli\u003eEbinger M, Galinovic I, Rozanski M, Brunecker P, Endres M, Fiebach JB. Fluid-attenuated inversion recovery evolution within 12 hours from stroke onset: a reliable tissue clock? Stroke. 2010 Feb;41(2):250\u0026ndash;5.\u003c/li\u003e\n \u003cli\u003eBerthez\u0026egrave;ne Y, Eker O, Makris N, Bettan M, Mansuy A, Chabrol A, et al. Collateral circulation assessment within the 4.5 h time window in patients with and without DWI/FLAIR MRI mismatch. J Neurol Sci. 2018 Nov;394:94\u0026ndash;8.\u003c/li\u003e\n \u003cli\u003eRimmele DL, Thomalla G. Wake-up stroke: clinical characteristics, imaging findings, and treatment option - an update. Front Neurol. 2014;5:35.\u003c/li\u003e\n \u003cli\u003eThomalla G, Cheng B, Ebinger M, Hao Q, Tourdias T, Wu O, et al. DWI-FLAIR mismatch for the identification of patients with acute ischaemic stroke within 4\u0026middot;5 h of symptom onset (PRE-FLAIR): a multicentre observational study. Lancet Neurol. 2011 Nov;10(11):978\u0026ndash;86.\u003c/li\u003e\n \u003cli\u003eBai Q-K, Zhao Z-G, Lu L-J, Shen J, Zhang J-Y, Sui H-J, et al. Treating ischaemic stroke with intravenous tPA beyond 4.5 hours under the guidance of a MRI DWI/T2WI mismatch was safe and effective. Stroke Vasc Neurol. 2019 Mar;4(1):8\u0026ndash;13.\u003c/li\u003e\n \u003cli\u003eThomalla G, Simonsen CZ, Boutitie F, Andersen G, Berthezene Y, Cheng B, et al. MRI-Guided Thrombolysis for Stroke with Unknown Time of Onset. N Engl J Med. 2018 Aug;379(7):611\u0026ndash;22.\u003c/li\u003e\n \u003cli\u003eJiang L, Peng M, Geng W, Chen H, Su H, Zhao B, et al. FLAIR hyperintensities-DWI mismatch in acute stroke: associations with DWI volume and functional outcome. Brain Imaging Behav. 2020 Aug;14(4):1230\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eWolman DN, van Ommen F, Tong E, Kauw F, Dankbaar JW, Bennink E, et al. Non-contrast dual-energy CT virtual ischemia maps accurately estimate ischemic core size in large-vessel occlusive stroke. Sci Rep. 2021 Mar;11(1):6745.\u003c/li\u003e\n \u003cli\u003eLiu L, Kurgan L, Wu F-X, Wang J. Attention convolutional neural network for accurate segmentation and quantification of lesions in ischemic stroke disease. Med Image Anal. 2020 Oct;65:101791.\u003c/li\u003e\n \u003cli\u003eCetinoglu YK, Koska IO, Uluc ME, Gelal MF. Detection and vascular territorial classification of stroke on diffusion-weighted MRI by deep learning. Eur J Radiol. 2021 Dec;145:110050.\u003c/li\u003e\n \u003cli\u003eBrinjikji W, Abbasi M, Arnold C, Benson JC, Braksick SA, Campeau N, et al. e-ASPECTS software improves interobserver agreement and accuracy of interpretation of aspects score. Interv Neuroradiol [Internet]. 2021;27(6):781\u0026ndash;7. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85104440160\u0026amp;doi=10.1177%2F15910199211011861\u0026amp;partnerID=40\u0026amp;md5=85700d836ed76ec5a818c6bc62dde819\u003c/li\u003e\n \u003cli\u003eWeyland CS, Papanagiotou P, Schmitt N, Joly O, Bellot P, Mokli Y, et al. Hyperdense Artery Sign in Patients With Acute Ischemic Stroke-Automated Detection With Artificial Intelligence-Driven Software. Front Neurol. 2022;13:807145.\u003c/li\u003e\n \u003cli\u003eRahman S, Sarker S, Haque AKMN, Uttsha MM, Islam MF, Deb S. AI-Driven Stroke Rehabilitation Systems and Assessment: A Systematic Review. IEEE Trans neural Syst Rehabil Eng a Publ IEEE Eng Med Biol Soc. 2023;31:192\u0026ndash;207.\u003c/li\u003e\n \u003cli\u003eMcInnes MDF, Moher D, Thombs BD, McGrath TA, Bossuyt PM, Clifford T, et al. Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies: The PRISMA-DTA Statement. JAMA. 2018 Jan;319(4):388\u0026ndash;96.\u003c/li\u003e\n \u003cli\u003eWhiting PF, Rutjes AWS, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011 Oct;155(8):529\u0026ndash;36.\u003c/li\u003e\n \u003cli\u003eBian R, Huo M, Liu W, Mansouri N, Tanglay O, Young I, et al. Connectomics underlying motor functional outcomes in the acute period following stroke. Front Aging Neurosci. 2023;15:1131415.\u003c/li\u003e\n \u003cli\u003eBonkhoff AK, Schirmer MD, Bretzner M, Etherton M, Donahue K, Tuozzo C, et al. Abnormal dynamic functional connectivity is linked to recovery after acute ischemic stroke. Hum Brain Mapp [Internet]. 2021;42(7):2278\u0026ndash;91. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85101879623\u0026amp;doi=10.1002%2Fhbm.25366\u0026amp;partnerID=40\u0026amp;md5=d925fcf7fbbb422add80eec0ab78c5f9\u003c/li\u003e\n \u003cli\u003eGandolfi M, Galazzo IB, Pavan RG, Cruciani F, Vale N, Picelli A, et al. eXplainable AI Allows Predicting Upper Limb Rehabilitation Outcomes in Sub-Acute Stroke Patients. IEEE J Biomed Heal Informatics [Internet]. 2023;27(1):263\u0026ndash;73. Available from: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85141606409\u0026amp;doi=10.1109%2FJBHI.2022.3220179\u0026amp;partnerID=40\u0026amp;md5=47174bb0378ea3178e20d18701ad7f96\u003c/li\u003e\n \u003cli\u003eChang H-H, Yeh S-J, Chiang M-C, Hsieh S-T. RU-Net: skull stripping in rat brain MR images after ischemic stroke with rat U-Net. BMC Med Imaging. 2023 Mar;23(1):44.\u003c/li\u003e\n \u003cli\u003eHuang S, Shen Q, Duong TQ. Artificial neural network prediction of ischemic tissue fate in acute stroke imaging. J Cereb Blood Flow Metab. 2010 Sep;30(9):1661\u0026ndash;70.\u003c/li\u003e\n \u003cli\u003eMoon HS, Heffron L, Mahzarnia A, Obeng-Gyasi B, Holbrook M, Badea CT, et al. Automated multimodal segmentation of acute ischemic stroke lesions on clinical MR images. Magn Reson Imaging. 2022 Oct;92:45\u0026ndash;57.\u003c/li\u003e\n \u003cli\u003eZhu HC, Jiang L, Zhang H, Luo LM, Chen Y, Chen YC. An automatic machine learning approach for ischemic stroke onset time identification based on DWI and FLAIR imaging. NEUROIMAGE-CLINICAL. 2021;31.\u003c/li\u003e\n \u003cli\u003eJiang L, Wang S, Ai Z, Shen T, Zhang H, Duan S, et al. Development and external validation of a stability machine learning model to identify wake-up stroke onset time from MRI. Eur Radiol. 2022 Jun;32(6):3661\u0026ndash;9.\u003c/li\u003e\n \u003cli\u003eAkay EMZ, Rieger J, Sch\u0026ouml;ttler R, Behland J, Schymczyk R, Khalil AA, et al. A deep learning analysis of stroke onset time prediction and comparison to DWI-FLAIR mismatch. NeuroImage Clin. 2023;40:103544.\u003c/li\u003e\n \u003cli\u003eYu H, Wang Z, Sun Y, Bo W, Duan K, Song C, et al. Prognosis of ischemic stroke predicted by machine learning based on multi-modal MRI radiomics. Front psychiatry. 2022;13:1105496.\u003c/li\u003e\n \u003cli\u003ePolson JS, Zhang H, Nael K, Salamon N, Yoo BY, El-Saden S, et al. Identifying acute ischemic stroke patients within the thrombolytic treatment window using deep learning. J Neuroimaging. 2022 Nov;32(6):1153\u0026ndash;60.\u003c/li\u003e\n \u003cli\u003eZhang H, Polson JS, Nael K, Salamon N, Yoo B, El-Saden S, et al. Intra-domain task-adaptive transfer learning to determine acute ischemic stroke onset time. Comput Med Imaging Graph. 2021 Jun;90:101926.\u003c/li\u003e\n \u003cli\u003eLee H, Lee E-J, Ham S, Lee H-B, Lee JS, Kwon SU, et al. Machine Learning Approach to Identify Stroke Within 4.5 Hours. Stroke. 2020 Mar;51(3):860\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eBenzakoun J, Deslys M-A, Legrand L, Hmeydia G, Turc G, Hassen W Ben, et al. Synthetic FLAIR as a Substitute for FLAIR Sequence in Acute Ischemic Stroke. Radiology. 2022 Apr;303(1):153\u0026ndash;9.\u003c/li\u003e\n \u003cli\u003eZhang R, Wang J. Machine Learning-Based Prediction of Subsequent Vascular Events After 6 Months in Chinese Patients with Minor Ischemic Stroke. Int J Gen Med. 2022;15:3797\u0026ndash;808.\u003c/li\u003e\n \u003cli\u003eLu J, Guo Y, Wang M, Luo Y, Zeng X, Miao X, et al. Determining acute ischemic stroke onset time using machine learning and radiomics features of infarct lesions and whole brain. Math Biosci Eng. 2024 Jan;21(1):34\u0026ndash;48.\u003c/li\u003e\n \u003cli\u003eCheng B, Pinnschmidt H, K\u0026ouml;nigsberg A, Schlemm E, Boutitie F, Ebinger M, et al. Estimating nocturnal stroke onset times by magnetic resonance imaging in the WAKE-UP trial. Int J stroke Off J Int Stroke Soc. 2022 Mar;17(3):323\u0026ndash;30.\u003c/li\u003e\n \u003cli\u003eJakubicek S, Krebs S, Posekany A, Ferrari J, Szabo J, Siarnik P, et al. Modified DWI-FLAIR mismatch guided thrombolysis in unknown onset stroke. J Thromb Thrombolysis. 2019 Feb;47(2):167\u0026ndash;73.\u003c/li\u003e\n \u003cli\u003eAdil MM, Luby M, Lynch JK, Hsia AW, Kalaria CP, Nadareishvili Z, et al. Routine use of FLAIR-negative MRI in the treatment of unknown onset stroke. J stroke Cerebrovasc Dis Off J Natl Stroke Assoc. 2020 Sep;29(9):105093.\u003c/li\u003e\n \u003cli\u003eBen Alaya I, Limam H, Kraiem T. Automatic triaging of acute ischemic stroke patients for reperfusion therapies using Artificial Intelligence methods and multiple MRI features: A review. Clin Imaging. 2023 Dec;104:109992.\u003c/li\u003e\n \u003cli\u003eGerbasi A, Konduri P, Tolhuisen M, Cavalcante F, Rinkel L, Kappelhof M, et al. Prognostic Value of Combined Radiomic Features from Follow-Up DWI and T2-FLAIR in Acute Ischemic Stroke. J Cardiovasc Dev Dis. 2022 Dec;9(12).\u003c/li\u003e\n \u003cli\u003eBen Alaya I, Limam H, Kraiem T. Applications of artificial intelligence for DWI and PWI data processing in acute ischemic stroke: Current practices and future directions. Clin Imaging. 2022 Jan;81:79\u0026ndash;86.\u003c/li\u003e\n \u003cli\u003eMiyamoto N, Ueno Y, Yamashiro K, Hira K, Kijima C, Kitora N, et al. Stroke classification and treatment support system artificial intelligence for usefulness of stroke diagnosis. Front Neurol. 2023;14:1295642.\u003c/li\u003e\n \u003cli\u003eLakatos L, Bolognese M, M\u0026uuml;ller M, \u0026Ouml;sterreich M, von Hessling A. Automated Supra- and Infratentorial Brain Infarct Volume Estimation on Diffusion Weighted Imaging Using the RAPID Software. Front Neurol. 2022;13:907151.\u003c/li\u003e\n \u003cli\u003eKrag CH, M\u0026uuml;ller FC, Gandrup KL, Raaschou H, Andersen MB, Brejneb\u0026oslash;l MW, et al. Diagnostic test accuracy study of a commercially available deep learning algorithm for ischemic lesion detection on brain MRIs in suspected stroke patients from a non-comprehensive stroke center. Eur J Radiol. 2023 Nov;168:111126.\u003c/li\u003e\n \u003cli\u003eFahed R, Lecler A, Sabben C, Khoury N, Ducroux C, Chalumeau V, et al. DWI-ASPECTS (Diffusion-Weighted Imaging-Alberta Stroke Program Early Computed Tomography Scores) and DWI-FLAIR (Diffusion-Weighted Imaging-Fluid Attenuated Inversion Recovery) Mismatch in Thrombectomy Candidates: An Intrarater and Interrater Agreement St. Stroke. 2018 Jan;49(1):223\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eScheldeman L, Wouters A, Dupont P, Christensen S, Boutitie F, Cheng B, et al. Diffusion-Weighted Imaging and Fluid-Attenuated Inversion Recovery Quantification to Predict Diffusion-Weighted Imaging-Fluid-Attenuated Inversion Recovery Mismatch Status in Ischemic Stroke With Unknown Onset. Stroke. 2022 May;53(5):1665\u0026ndash;73.\u003c/li\u003e\n \u003cli\u003eCivrny J, Sedlackova Z, Malenak T, Kucera P, Machal D, Kocher M, et al. Comparison of semi-quantitative and visual assessment of early MRI signal evolution in acute ischaemic stroke. Eur J Radiol open. 2023;10:100488.\u003c/li\u003e\n \u003cli\u003eGalinovic I, Dicken V, Heitz J, Klein J, Puig J, Guibernau J, et al. Homogeneous application of imaging criteria in a multicenter trial supported by investigator training: A report from the WAKE-UP study. Eur J Radiol. 2018 Jul;104:115\u0026ndash;9.\u003c/li\u003e\n \u003cli\u003eRegenhardt RW, Bretzner M, Zanon Zotin MC, Bonkhoff AK, Etherton MR, Hong S, et al. Radiomic signature of DWI-FLAIR mismatch in large vessel occlusion stroke. J neuroimaging Off J Am Soc Neuroimaging. 2022 Jan;32(1):63\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eGirot M, Leclerc X, Gauvrit J-Y, Verdelho A, Pruvo J-P, Leys D. Cerebral magnetic resonance imaging within 6 hours of stroke onset: inter- and intra-observer reproducibility. Cerebrovasc Dis. 2003;16(2):122\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eKits A, Al-Saadi J, De Luca F, Janzon F, Mazya M V, Lundberg J, et al. 2.5-Minute Fast Brain MRI with Multiple Contrasts in Acute Ischemic Stroke. Neuroradiology. 2024 Mar;\u003c/li\u003e\n \u003cli\u003eVerclytte S, Gnanih R, Verdun S, Feiweier T, Clifford B, Ambarki K, et al. Ultrafast MRI using deep learning echoplanar imaging for a comprehensive assessment of acute ischemic stroke. Eur Radiol. 2023 May;33(5):3715\u0026ndash;25.\u003c/li\u003e\n \u003cli\u003eHo KC, Speier W, Zhang H, Scalzo F, El-Saden S, Arnold CW. A Machine Learning Approach for Classifying Ischemic Stroke Onset Time From Imaging. IEEE Trans Med Imaging. 2019 Jul;38(7):1666\u0026ndash;76.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 3 are available in the Supplementary Files section\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, DWI/FLAIR mismatch, stroke, dice similarity coefficient, MRI. ","lastPublishedDoi":"10.21203/rs.3.rs-4494281/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4494281/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe purpose of this study was to survey the existing artificial intelligence (AI) algorithms created for the automated detection of the diffusion-weighted imaging (DWI)\u0026ndash;fluid-attenuated inversion recovery (FLAIR) mismatch and assess how their performance compares to that diagnostic techniques performed by neuroradiologists. The literature search for this systematic review was conducted in PubMed, MEDLINE, Ovid Embase, Web of Science, Scopus, and Cochrane databases up until February 2, 2024. The review team cross-checked the reference lists of the included studies to identify any additional relevant references, in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. We assessed the included studies using the Quality Assessment of Diagnostic Accuracy Studies 2 tool. The area under the curve (AUC) was reported in most studies, However, one study did not report this metric, The AI models achieved AUCs between 0.60 and 0.90, Sensitivity ranged from 0.6 to 0.9, and specificity ranged from 0.72 to 0.99, the positive predictive value (PPV), negative predictive value (NPV) and F1-Score were ranging from 0.72 to 0.93, 0.47 to 0.91 and 0.65 to 0.9, respectively. Additionally, the dice similarity coefficients (DSC) 0.73 \u0026amp; 0.8 were stated in two researches and accuracies ranging from 0.67 to 0.99. This review indicates that the current AI methods for DWI/FLAIR mismatch assessment may not be able to accurately determine the time since stroke onset based only on the DWI and FLAIR sequences, however, an AI-based approach focused on treatment eligibility, outcome prediction, and incorporating patient-specific information could potentially improve the care of stroke patients.\u003c/p\u003e","manuscriptTitle":"A systematic review of the performance of Artificial Intelligence for automated DWI/FLAIR mismatch evaluation on MRI in ischemic stroke","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-19 15:12:10","doi":"10.21203/rs.3.rs-4494281/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"dce93f84-f277-4a3d-880c-e32dd6c035a4","owner":[],"postedDate":"June 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-06-19T15:12:10+00:00","versionOfRecord":[],"versionCreatedAt":"2024-06-19 15:12:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4494281","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4494281","identity":"rs-4494281","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}