Deep-learning–based non-contrast CT for detecting acute ischemic stroke: a systematic review and HSROC meta-analysis of patient-level diagnostic accuracy Authors: Kalab Yigermal Gete, Asnakew Achaw Ayele Published: 2025 ABSTRACT AbstractBackgroundNon-contrast CT (NCCT) is first-line imaging for suspected acute ischemic stroke (AIS) but has limited early sensitivity; deep learning (DL) may improve patient-level detection.ObjectivesTo estimate the diagnostic accuracy of DL applied to NCCT for patient-level AIS detection and to examine prespecified sources of between-study heterogeneity.MethodsWe searched MEDLINE, Embase, and Web of Science (January 2010–May 2025). Eligible prospective or retrospective diagnostic studies evaluated DL on NCCT against an appropriate reference standard and reported (or allowed reconstruction of) patient-level 2 × 2 data. Two-gate case–control and lesion-only reports were excluded. Dual reviewers screened/extracted data; risk of bias was assessed with QUADAS-2, and AI-reporting against items adapted from STARD-AI/CLAIM/CONSORT-AI. Bivariate random-effects/HSROC models summarized sensitivity and specificity. Prespecified moderators were posterior-fossa inclusion, reference-standard robustness, and validation type. Sensitivity analyses included external-only cohorts, robust standards, posterior-fossa inclusion, and a “Direct AIS” construct subset.ResultsOf 1,899 records, 16 studies met inclusion; 13 contributed patient-level data to meta-analysis. Summary sensitivity was 0.91 (95% CI, 0.81–0.96) and specificity 0.90 (0.85–0.94). Sensitivity was lower for externally validated models than internally validated ones (0.82 [0.67–0.91] vs. 0.95 [0.89–0.98]) with similar specificity (0.88 [0.83–0.92] vs. 0.93 [0.82–0.97]). Findings were directionally robust across sensitivity analyses. QUADAS-2 frequently indicated concerns in patient selection and index-test domains; AI-reporting quality was mostly moderate, and explicit external validation remained uncommon.ConclusionsDL applied to NCCT shows high accuracy for patient-level AIS detection. However, generalizability is the principal gap; broader external validation and guideline-concordant reporting are needed to support safe clinical adoption.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12883-025-04528-3.Keywords: Deep learning, Non-contrast computed tomography (NCCT), Acute ischemic stroke, Diagnostic test accuracy, Sensitivity and specificity, Hierarchical summary ROC (HSROC), Patient-level meta-analysis, Systematic review Introduction Acute ischemic stroke (AIS) is a leading cause of death and disability worldwide, and outcomes depend critically on rapid assessment and time-sensitive reperfusion therapies within organized systems of care [1]. In contemporary emergency pathways, non-contrast CT (NCCT) is the first-line imaging test for suspected stroke because it is fast, broadly available, and reliably excludes intracranial hemorrhage while offering an initial view of early ischemic change (EIC) [1]. However, for ischemia detection per se, NCCT is substantially less sensitive than diffusion-weighted MRI (DWI), particularly early after symptom onset; in a landmark prospective comparison, MRI identified acute ischemia far more often than CT across the emergency stroke population [2, 3]. These realities motivate methods that can enhance patient-level AIS detection on NCCT without slowing triage. The diagnostic challenge is not uniform across AIS presentations. Sensitivity of NCCT is especially limited in the posterior circulation, where beam-hardening from the skull base, small lesion volumes, and slower radiographic evolution obscure EIC—factors that matter at the bedside [4, 5]. Even in the anterior circulation, the visibility of EIC depends on time-to-CT and infarct burden; very early scans (≤ 3–6 h) often show subtle gray–white blurring or loss of the insular ribbon that can be missed under routine windowing [2]. These disease- and physics-specific issues suggest that any technology designed to improve NCCT detection should be evaluated with attention to territory (posterior fossa inclusion), onset-to-scan times, and infarct extent. It is also essential to distinguish three distinct tasks on NCCT that are frequently conflated in the literature. First, direct AIS detection is a patient-level classification (“AIS present/absent on NCCT”) intended to reduce misses and accelerate care. Second, surrogate vascular signs such as the hyperdense artery sign (HAS) indicate large-vessel occlusion (LVO) risk rather than ischemic parenchyma itself; meta-analytic data show HAS is moderately accurate for arterial obstruction, but it is not a parenchymal AIS detector and its prevalence varies with timing and technique [6, 7]. Third, severity/topography scoring with the Alberta Stroke Program Early CT Score (ASPECTS) quantifies EIC in the MCA territory for prognostication and treatment selection; ASPECTS is complementary to, but conceptually distinct from, binary AIS detection [8, 9]. Conflating these constructs—e.g., pooling ASPECTS estimation, HAS detection, slice-level lesion finding, and patient-level AIS diagnosis—muddies clinical interpretability and can inflate performance estimates. Deep-learning (DL) systems promise to enhance NCCT interpretation by amplifying subtle EIC cues that humans find difficult early after onset. Proof-of-concept studies using MRI-DWI as reference standard demonstrate that DL can localize and delineate early infarction on NCCT, explicitly addressing the modality’s well-known insensitivity in the first hours [10]. At the same time, the broader AI-stroke literature remains heterogeneous in tasks (detection vs. ASPECTS vs. surrogates), reference standards (e.g., clinical follow-up vs. DWI), validation (internal vs. external), and operating-point selection—all of which shape measured sensitivity and specificity and hence transportability to real-world triage. Recent narrative and systematic reviews highlight growing capability but also emphasize the need for standardized reporting and external validation before clinical deployment, particularly for ischemic-stroke NCCT tasks where heterogeneity and generalizability remain key challenges [11]. These gaps make a disease-specific, construct-clean synthesis—focused on patient-level AIS detection on NCCT—both timely and clinically meaningful. Accordingly, this systematic review and patient-level meta-analysis aims to provide robust, clinically interpretable estimates of diagnostic accuracy for DL applied to NCCT in suspected AIS at presentation. We prespecify moderators that reflect AIS pathophysiology and NCCT physics—not generic AI concerns—including time-to-CT, infarct burden (ASPECTS), and posterior fossa inclusion, alongside design features known to affect transportability (reference-standard robustness, validation type, and operating-point choice; DL used alone vs. as an assist). Locking the unit of analysis to the patient improves clinical interpretability and supports translation into emergency triage workflows. Methods and reporting follow PRISMA-DTA [12] with QUADAS-2 [13] risk-of-bias assessment; where applicable, we map AI-specific reporting items to CONSORT-AI [14] and CLAIM [15] to enhance transparency and reproducibility. Methods Eligibility criteria We included prospective or retrospective cross-sectional and cohort studies evaluating the diagnostic accuracy of DL models applied to NCCT for detecting AIS. Eligible studies required both the index test (standalone DL or DL-assist) and a valid reference standard, with sufficient information to construct patient-level 2 × 2 tables (TP, FP, FN, TN). We excluded two-gate case–control designs with non-representative controls, case reports, reviews, editorials, abstracts without full data, unverifiable accuracy data, and duplicates (retaining the most complete version). Lesion/image-level results were retained for narrative context but excluded from the patient-level meta-analysis. Participants Adults (≥ 18 years) undergoing NCCT for suspected AIS in emergency, inpatient, or acute outpatient settings were eligible. We excluded paediatric cohorts, hemorrhage-only cohorts lacking an AIS subgroup, non-acute indications only, and series where baseline scans could not be separated from post-reperfusion imaging. Index test The index tests were DL systems applied to NCCT yielding patient-level decisions (AIS present/absent), used as standalone DL or DL-assist. Any DL architecture and typical NCCT pre/post-processing were eligible. We excluded traditional ML/radiomics without DL, non-NCCT pipelines (e.g., CTA/CTP-only), or studies without a clear diagnostic operating point. Target condition AIS present at the time of the index NCCT. Subtypes (e.g., anterior/posterior, LVO/non-LVO, infarct size) were eligible if analyzable at the patient level. TIA without objective evidence of acute ischemia (e.g., DWI-positive lesion) was excluded. Reference standard Preferred standards were MRI-DWI for confirmation of AIS, with negatives verified by MRI or adequate clinical/imaging follow-up. We also accepted a clearly defined composite reference standard based on blinded expert consensus using all available clinical and imaging information; when MRI/follow-up were unavailable, we accepted expert interpretation independent of the index algorithm (single- or multi-reader, with or without consensus), coded as ‘mixed/weaker’ in moderator analyses. Data and language restrictions Human studies from 2010 onward were eligible. Non-English studies were included only if an accessible English translation was available. Studies had to report—or allow reconstruction of—sensitivity, specificity, or full patient-level 2 × 2 data. Information sources and search strategy We searched MEDLINE (Ovid), Embase (Ovid), and Web of Science (January 2010 to May 2025), combining controlled vocabulary and free-text for DL, AIS, NCCT, and diagnostic accuracy. Full strategies are provided in the Supplement (Tables S1–S3). Study selection Rayyan was used for deduplication and screening. Two reviewers independently screened titles/abstracts and full texts against prespecified criteria; disagreements were resolved by discussion or third-party adjudication. Reasons for exclusion were recorded. Data extraction Two reviewers independently extracted data using a piloted, standardised form. Items included study design/recruitment, participants and setting, index-test details (architecture; standalone vs. DL-assist; operating point; validation type), reference standards, flow/timing, and patient-level performance. One record was captured per unique patient-level arm (DL-alone; DL-assist). Consistency checks verified TP + FN and TN + FP totals, prevalence, and analysed N. Where only sensitivity/specificity with denominators or prevalence were available, we reconstructed 2 × 2 values with documented assumptions; authors were contacted as needed. Lesion-level results were recorded separately and not pooled. When multiple candidate arms were reported, one primary patient-level estimate per track (DL-alone; DL-assist) was selected using a predefined hierarchy: external over internal test sets; pre-specified/locked over post-hoc thresholds; representative cohorts (ED all-comers; all territories) over enriched subsets; robust reference standards over mixed/weaker; multicentre over single-centre; and vendor-locked over statistically optimised operating points. All choices/ties were logged. Data definitions The index test was DL on NCCT (standalone or DL-assist). The target condition was AIS at index presentation. Sensitivity and specificity were TP/(TP + FN) and TN/(TN + FP). Prespecified covariates for subgroup/meta-regression were: time-to-CT (≤ 6 h vs. > 6–24 h/mixed), ASPECTS (≤ 7 vs. > 7), posterior fossa included (yes vs. anterior-only), reference-standard robustness (robust vs. mixed/weaker), validation type (external vs. internal), operating-point selection (pre-specified/locked vs. post-hoc optimised), and mode of use (DL-assist vs. DL-alone). Coding used majority rule (≥ 60%), fixed cut-points, and documented sources. Risk of bias and reporting quality Two reviewers assessed risk of bias with QUADAS-2 (Patient Selection, Index Test, Reference Standard, Flow/Timing); disagreements were resolved by consensus or third reviewer. AI-specific reporting was evaluated across seven items adapted from STARD-AI, CLAIM, and CONSORT-AI—data source, training/test split, external validation, explainability, model availability, workflow integration, and regulatory status—and summarised; studies were classified as Good, Moderate, or Poor using a predefined rubric. Statistical analysis Primary outcomes were pooled sensitivity and specificity. We fitted hierarchical summary ROC (HSROC) and bivariate random-effects (Reitsma) models, selecting one operating point per study by hierarchy; independently tested, non-overlapping arms were treated as separate studies. Proportions were analysed on the logit scale with small continuity corrections for zero cells. Outputs included paired forest plots and an HSROC curve with summary point and 95% prediction region; diagnostic odds and likelihood ratios were derived when available. Subgroup analyses and univariable meta-regression were performed for prespecified covariates when both strata contained sufficient studies (≥ 4). Analyses used Stata v17 and R v4.5.1. Additional Analyses Sensitivity analyses restricted to (i) external validation only, (ii) robust reference standards only (DWI-confirmed positives with adequately verified negatives), (iii) cohorts including the posterior fossa (excluding anterior-only studies), and (iv) a Direct AIS construct subset (patient-level AIS detection; excludes CTA-defined LVO/HAS and ASPECTS-only tools). Results Study selection A systematic search of MEDLINE, Embase, and Web of Science (January 2010–May 2025) yielded 1,899 records. After removing 390 duplicates, 1,509 titles/abstracts were screened, of which 1,403 were excluded as irrelevant (e.g., non-diagnostic focus, conventional CAD/non-DL, non-NCCT, lesion-only reports, conference abstracts, or insufficient outcome data). Full texts of 106 articles were assessed; 90 were excluded (common reasons: two-gate case–control designs, absence of patient-level outcomes or a reconstructible 2 × 2 table, reliance on non-robust/inadequately verified reference standards, or index tests without a clear operating point). Sixteen studies met inclusion; 13 provided patient-level 2 × 2 data and were pooled; 3 reported lesion/image-level accuracy only and were summarised narratively (not pooled). PRISMA flow is shown in Fig. 1. PRISMA 2020 flow diagram Sixteen studies (15 retrospective, 1 prospective) spanned Asia, Europe, and North America, evenly split between single-centre (8/16) and multicentre (8/16) cohorts. Adults with suspected acute stroke underwent NCCT; several focused on hyperacute windows, LVO/HAS, or ASPECTS-defined severity. Thirteen studies reported patient-level AIS detection and entered the meta-analysis; three were lesion/image-level and were summarised descriptively. Patient-level sample sizes ranged widely (≈ 25–864). Index tests included commercial tools (qER/Qure.ai, Brainomix e-Stroke/e-ASPECTS, RAPID NCCT Stroke) and research CNN/3D-U-Net pipelines. Reference standards were most often MRI-DWI (with clinical follow-up), or expert consensus where MRI was unavailable (Table 1). A test-level covariate matrix is provided in the Supplementary Materials, Table S4. Table 1Summary of study characteristicsStudy IDDesign; Centers; CountryInclusion criteria; Exclusion criteriaUnit of analysis; Sample size (n)Index test nameTarget conditionReference standard Babutain 2021 [16]Retrospective; Single; Saudi ArabiaAdults with suspected acute stroke undergoing index NCCT;Poor-quality/artefact NCCT; hemorrhagic stroke on index CT.Patient-level; n = 25DL classifier for AIS detection on NCCTAIS present at the time of index NCCTExpert CT interpretation by medical team Chen 2022 [17]Retrospective; Single; TaiwanUnenhanced brain NCCT images labeled normal/hemorrhage/infarct/other;Not explicitly statedImage-level; n = 2,476DL classifiers (CNN-2, VGG-16, ResNet-50; transfer learning) — multi-class NCCT categorization.Infarct class on NCCT within a multi-class image-level taskClinical labels with follow-up and expert radiologist review. Chiramal 2024 [18]Retrospective; Single; IndiaConsecutive NCCT scans for suspected stroke with both AI output and teleradiology reports;Missing AI or missing teleradiology report.Patient-level; n=864qER (Qure.ai) — automated NCCT brain interpretation.AIS (infarct) present at the time of index NCCTTeleradiologist report used as ground truth Dyer 2021 [19]Retrospective; Multi; UK, USA, and IndiaNCCT head scans with FRCR radiologist ground truth;Corrupted/incomplete scansPatient-level; n = 390EfficientNet-based DL model — NCCT classification ± weak segmentation.AIS present at the time of index NCCTExpert radiologist ground truth on CT Fuchigami 2020 [20]Retrospective; Single; JapanAdults with hyperacute ischemic stroke (≤ 4.5 h) undergoing NCCT;Not statedPatient-level; n=703D U-Net with left–right comparator — hyperacute ischemic lesion segmentation on NCCTAIS present at the time of index NCCT in the hyperacute windowMRI-DWI ± FLAIR within ~ 1 h Gauriau 2023 [10]Retrospective; Multi; USAAdmission NCCT for suspected AIS with MRI-DWI ≤ 3 h (positives)/≤5 days (negatives); earliest NCCT used;Non-diagnostic/artefact CT; failed/imprecise registrations.Patient-level; n=3643D U-Net–based DL model — early infarct detection/segmentation on NCCT.AIS present at the time of index NCCTMRI-DWI ± ADC with clinical correlation; positives verified ≤ 3 h, negatives within ≤ 5 days. Herweh 2016 [21]Retrospective; Single; GermanyAcute stroke patients with CT and DWI < 2 h apart; suspected unilateral anterior circulation ischemia;Poor-quality/artefact CT.Lesion-level; n=680e-ASPECTS — automated ASPECTS scoring on NCCT.Ischemic severity on NCCT (ASPECTS)MRI-DWI consensus Lo 2021 [22]Retrospective; Multi; TaiwanAdults with AIS ≤ 6 h and ICA/M1 occlusion on CTA;Vertebrobasilar occlusion; arrival > 6 h; motion artefactsLesion-level; n=225Slice-level DCNNs (AlexNet, Inception-v3, ResNet-101; transfer) — NCCT classification.AIS—LVO subtype (arterial occlusion surrogate) on NCCTMultiphase CTA (occlusion presence and site). Lu 2022 [23]Retrospective; Multi; ChinaSuspected AIS with NCCT ≤ 24 h and MRI (incl. DWI) ≤ 72 h;Head injury/brain tumor; NCCT–DWI mismatch; strong CT artefacts.Patient-level; n=150Two-stage DL model — YOLOv3 (localization) + ResNet-50 (classification) for early/invisible AIS on NCCT.AIS present at the time of index NCCTMRI-DWI plus clinical diagnosis. Mallon 2023 [24]Prospective; Single; UKConsecutive acute stroke admissions with NCCT (CTA/CTP per pathway);No index NCCT; mass lesion on initial CTPatient-level; n=492Brainomix e-Stroke (e-ASPECTS) — automated ASPECTS analysis on NCCT.AIS—severity subtype on NCCT using ASPECTS-based patient-level thresholdingExpert neuroradiologist ground truth using all imaging and clinical data Pastor-Vargas 2025 [25]Retrospective; Single; SpainNCCT for suspected ischemia;Hemorrhage or non-ischemic conditions on NCCT.Patient-level; n=316DenseNet-3D CNN — ischemia detection on NCCT (transfer learning).AIS present at the time of index NCCTMRI-DWI with clinical follow-up. Sahoo 2022 [26]Retrospective; Single; TaiwanAdults with ischemic stroke onset < 12 h;DWI infarct < 0.5 cm; traumatic brain injury; brain malignancy; ICHPatient-level; n=58Customized VGG-16 CNN — ischemic lesion identification on NCCT (2D slices).AIS present at the time of index NCCTMRI-DWI Wei 2025 [27]Retrospective; Multi; ChinaAdults with suspected AIS, last-known-well ≤ 8 h, undergoing baseline NCCT (ASPECTS tool);Intracranial hemorrhage; prior craniotomy; traumatic ICH; poor-quality images precluding diagnosisPatient-level; n=220DL-based automated ASPECTS system — 3D ResNet (localize) + 2.5D U-Net (segment) + 2.5D ResNet (score).AIS—severity subtype on NCCT using ASPECTS-based patient-level thresholdingExpert consensus ASPECTS on NCCT (± CTP if available) Weyland 2022 [7]Retrospective; Multi; Germany, Greece, and UKAcute stroke or stroke-like symptoms with both NCCT and CTA available (HAS/LVO task);Not explicitly statedPatient-level; n=154Brainomix HAS detector — AI-driven hyperdense artery sign detection on NCCT.AIS—LVO subtype on NCCTCTA for LVO presence and side. Wu 2021 [28]Retrospective; Multi; ChinaStroke cases with NCCT and DWI (onset-to-NCCT < 24 h);Not explicitly statedPatient-level; n=77Two-stage CNN — U-Net (segmentation) + ResNet (patch classifier) with MAP post-processing on NCCT.AIS present at the time of index NCCTMRI-DWI for positives ± clinical confirmation Yedavalli 2023 [29]Retrospective; Multi; USAAcutely presenting suspected stroke with index NCCT; LVO defined (distal intracranial ICA and/or M1 MCA);Not explicitly detailed.Patient-level; n=244RAPID NCCT Stroke — automated NCCT analysis platform.AIS—LVO subtype within suspected AIS cohortCTA for LVO with consensus of two of three neuroradiologists.NCCT non-contrast, CT AIS acute ischemic stroke;, LVO large-vessel occlusion, HAS hyperdense artery sign, CTA CT angiography, MRI-DWI/ADC diffusion-weighted MRI/apparent diffusion coefficient, ASPECTS Alberta Stroke Program Early CT Score, CTP CT perfusion, MCA middle cerebral artery. Unit of analysis rows marked patient-level contributed to the primary HSROC meta-analysis, image- or lesion-level rows are descriptive only (not pooled). Moderators test-level covariates (time-to-CT window, ASPECTS burden, posterior fossa inclusion, RS robustness, validation type, operating-point selection, mode of use) are encoded in the supplementary matrix with codes 0/1/2 as pre-specified. Vendor/platform names appear as reported; custom models are summarized by backbone + task Summary of study characteristics Adults with suspected acute stroke undergoing index NCCT; Poor-quality/artefact NCCT; hemorrhagic stroke on index CT. Unenhanced brain NCCT images labeled normal/hemorrhage/infarct/other; Not explicitly stated Consecutive NCCT scans for suspected stroke with both AI output and teleradiology reports; Missing AI or missing teleradiology report. Patient-level; n= 864 NCCT head scans with FRCR radiologist ground truth; Corrupted/incomplete scans Adults with hyperacute ischemic stroke (≤ 4.5 h) undergoing NCCT; Not stated Patient-level; n= 70 Admission NCCT for suspected AIS with MRI-DWI ≤ 3 h (positives)/≤5 days (negatives); earliest NCCT used; Non-diagnostic/artefact CT; failed/imprecise registrations. Patient-level; n= 364 Acute stroke patients with CT and DWI < 2 h apart; suspected unilateral anterior circulation ischemia; Poor-quality/artefact CT. Lesion-level; n= 680 Adults with AIS ≤ 6 h and ICA/M1 occlusion on CTA; Vertebrobasilar occlusion; arrival > 6 h; motion artefacts Lesion-level; n= 225 Suspected AIS with NCCT ≤ 24 h and MRI (incl. DWI) ≤ 72 h; Head injury/brain tumor; NCCT–DWI mismatch; strong CT artefacts. Patient-level; n= 150 Consecutive acute stroke admissions with NCCT (CTA/CTP per pathway); No index NCCT; mass lesion on initial CT Patient-level; n= 492 NCCT for suspected ischemia; Hemorrhage or non-ischemic conditions on NCCT. Patient-level; n= 316 Adults with ischemic stroke onset < 12 h; DWI infarct < 0.5 cm; traumatic brain injury; brain malignancy; ICH Patient-level; n= 58 Adults with suspected AIS, last-known-well ≤ 8 h, undergoing baseline NCCT (ASPECTS tool); Intracranial hemorrhage; prior craniotomy; traumatic ICH; poor-quality images precluding diagnosis Patient-level; n= 220 Acute stroke or stroke-like symptoms with both NCCT and CTA available (HAS/LVO task); Not explicitly stated Patient-level; n= 154 Stroke cases with NCCT and DWI (onset-to-NCCT < 24 h); Not explicitly stated Patient-level; n= 77 Acutely presenting suspected stroke with index NCCT; LVO defined (distal intracranial ICA and/or M1 MCA); Not explicitly detailed. Patient-level; n= 244 NCCT non-contrast, CT AIS acute ischemic stroke;, LVO large-vessel occlusion, HAS hyperdense artery sign, CTA CT angiography, MRI-DWI/ADC diffusion-weighted MRI/apparent diffusion coefficient, ASPECTS Alberta Stroke Program Early CT Score, CTP CT perfusion, MCA middle cerebral artery. Unit of analysis rows marked patient-level contributed to the primary HSROC meta-analysis, image- or lesion-level rows are descriptive only (not pooled). Moderators test-level covariates (time-to-CT window, ASPECTS burden, posterior fossa inclusion, RS robustness, validation type, operating-point selection, mode of use) are encoded in the supplementary matrix with codes 0/1/2 as pre-specified. Vendor/platform names appear as reported; custom models are summarized by backbone + task Risk of bias and reporting quality QUADAS-2 Across 16 studies overall, risk-of-bias concerns peaked in Patient Selection (High 10/16, 62.5%; Unclear 4/16, 25.0%; Low 2/16, 12.5%) and Index Test (Unclear 8/16, 50.0%; High 4/16, 25.0%; Low 4/16, 25.0%). Reference Standard was mostly Unclear (13/16, 81.2%) with Low 3/16 (18.8%) and no High. Flow/Timing had the highest Low proportion (6/16, 37.5%) yet notable High (5/16, 31.2%) and Unclear (5/16, 31.2%). Applicability was largely acceptable for Index Test (Low 15/16, 93.8%) and Reference Standard (Low 11/16, 68.8%), whereas Patient Selection often raised concerns (High 11/16, 68.8%). (Fig. 2) Visual summary of the QUADAS-2 assessment AI reporting (STARD-AI/CLAIM items) Data source was described in 16/16 (100%) studies. Training/test splits were reported in 12/16 (75.0%) (Unclear 1/16, 6.3%; Not reported 3/16, 18.8%). External validation appeared in 3/16 (18.8%) (Unclear 7/16, 43.8%; Not reported 6/16, 37.5%). Explainability was reported in 7/16 (43.8%) (Unclear 1/16, 6.3%; Not reported 8/16, 50.0%). Model availability (code/weights or named commercial system) was reported in 6/16 (37.5%); workflow integration in 10/16 (62.5%) (Unclear 4/16, 25.0%); and regulatory status in 3/16 (18.8%). Overall, 1/16 (6.3%) study met the Good rubric, 11/16 (68.8%) were Moderate, and 4/16 (25.0%) were Poor (Table 2). Table 2AI-specific reporting quality assessmentStudy IDData Source DescribedTraining/Test Split ReportedExternal ValidationModel ExplainabilityModel AvailabilityClinical Workflow IntegrationRegulatory Status DiscussedOverall Assessment Babutain 2021 [16]ReportedReportedNot reportedNot reportedNot reportedUnclearNot reportedPoor Chen 2022 [17]ReportedReportedNot reportedNot reportedNot reportedReportedNot reportedPoor Chiramal 2024 [18]ReportedNot reportedUnclearReportedReportedReportedReportedModerate Dyer 2021 [19]ReportedReportedUnclearReportedNot reportedReportedNot reportedModerate Fuchigami 2020 [20]ReportedReportedNot reportedReportedNot reportedNot reportedNot reportedPoor Gauriau 2023 [10]ReportedReportedNot reportedReportedNot reportedReportedNot reportedModerate Herweh 2016 [21]ReportedReportedUnclearNot reportedReportedReportedReportedModerate Lo 2021 [22]ReportedReportedReportedNot reportedNot reportedUnclearNot reportedModerate Lu 2022 [23]ReportedReportedReportedReportedReportedReportedNot reportedGood Mallon 2023 [24]ReportedNot reportedUnclearUnclearReportedReportedNot reportedModerate Pastor-Vargas 2025 [25]ReportedReportedNot reportedNot reportedNot reportedUnclearNot reportedPoor Sahoo 2022 [26]ReportedReportedNot reportedNot reportedNot reportedReportedNot reportedModerate Wei 2025 [27]ReportedReportedUnclearReportedNot reportedReportedNot reportedModerate Weyland 2022 [7]ReportedUnclearUnclearReportedReportedUnclearNot reportedModerate Wu 2021 [28]ReportedReportedReportedNot reportedNot reportedNot reportedNot reportedModerate Yedavalli 2023 [29]ReportedNot reportedUnclearNot reportedReportedReportedReportedModerateSeven AI reporting items adapted from STARD-AI/CLAIM/CONSORT-AI were evaluated: (1) data source described; (2) training/test split reported; (3) external validation on an independent dataset; (4) model explainability (e.g., saliency/attribution); (5) model availability (code/weights or named commercial system); (6) clinical workflow integration (mode of use, reader-study/deployment details); and (7) regulatory status discussed. Each item was graded as reported, Unclear, or not reported. Overall categories: Good = all three core items (training/test split, external validation, explainability) reported and ≥ 5/7 items reported (with “Unclear” not counting toward the total and ≤ 1 other item “Not reported”); Moderate = 3–4/7 items reported, or ≥ 5/7 reported with ≥ 1 core item Unclear/Not reported; Poor = ≤ 2/7 items reported, or ≥ 4 items not reported, or both training/test split and external validation not reported AI-specific reporting quality assessment Seven AI reporting items adapted from STARD-AI/CLAIM/CONSORT-AI were evaluated: (1) data source described; (2) training/test split reported; (3) external validation on an independent dataset; (4) model explainability (e.g., saliency/attribution); (5) model availability (code/weights or named commercial system); (6) clinical workflow integration (mode of use, reader-study/deployment details); and (7) regulatory status discussed. Each item was graded as reported, Unclear, or not reported. Overall categories: Good = all three core items (training/test split, external validation, explainability) reported and ≥ 5/7 items reported (with “Unclear” not counting toward the total and ≤ 1 other item “Not reported”); Moderate = 3–4/7 items reported, or ≥ 5/7 reported with ≥ 1 core item Unclear/Not reported; Poor = ≤ 2/7 items reported, or ≥ 4 items not reported, or both training/test split and external validation not reported Results of individual studies Across the 13 pooled arms, study-level sensitivity ranged from 58.6% to 100.0% (median 88.6%, IQR 79.5–95.4%) and specificity from 63.8% to 98.0% (median 84.0%, IQR 77.2–90.3%). Individual estimates with 95% CIs are shown in Fig. 3. Coupled forest plots of sensitivity and specificity Meta-analysis Base HSROC model Across the 13 pooled arms, the bivariate/HSROC summary point was sensitivity 0.91 (95% CI 0.81–0.96) and specificity 0.90 (0.85–0.94). Overall discrimination was strong (see DOR and likelihood ratios in Table 3). Heterogeneity was greater for specificity than sensitivity; the positive Se–Sp correlation (0.41; 95% CI − 0.27–0.82) suggests between-study heterogeneity was not dominated by simple threshold shifts; rather, cohorts with clearer signal tended to improve in both metrics. HSROC parameters were λ = 4.59 (95% CI 3.41–5.77), θ = −0.44 (− 1.24 to 0.36), β = −0.42 (− 1.15 to 0.32; not different from 0), with between-study variances s²α = 3.05 (1.10–8.49) and s²θ = 0.32 (0.11–0.94) (Fig. 4) Prespecified covariates eligible for both subgroup analysis and univariable meta-regression were: posterior fossa (excluded vs. included), reference-standard robustness (mixed/weaker vs. robust), and validation type (internal vs. external). Table 3Summary of pooled estimatesSubgroupStudies (n)Sensitivity (95% CI)Specificity (95% CI)DOR (95% CI)LR+ (95% CI)LR– (95% CI)Overall130.91 (0.81–0.96)0.90 (0.85–0.94)90.29 (29.32–278.04)9.27 (5.64–15.24)0.10 (0.05–0.23)Posterior fossa IncludedYes (Included)70.95 (0.83–0.98)0.88 (0.78–0.93)124.22 (24.64–626.22)7.65 (4.05–14.46)0.06 (0.02–0.21)Anterior only (Excluded)40.77 (0.53–0.91)0.93 (0.85–0.97)41.66 (7.39–234.78)10.48 (4.04–27.22)0.25 (0.10–0.60)Reference standard robustnessRobust (DWI-MRI for positives; negatives verified by MRI or adequate follow-up)60.90 (0.76–0.97)0.89 (0.78–0.94)73.53 (19.82–272.77)7.90 (4.02–15.52)0.11 (0.04 − 0.29)Mixed/Weaker70.92 (0.73–0.98)0.92 (0.85–0.96)133.80 (18.38–973.95)11.46 (5.51–23.83)0.09 (0.02–0.34)Validation typeExternal60.82 (0.67–0.91)0.88 (0.83–0.92)33.37 (14.32–77.78)6.95 (4.60–10.49)0.21 (0.11–0.39)Internal60.95 (0.89–0.98)0.93 (0.82–0.97)258.36 (41.20–1620.32)13.29 (4.87–36.27)0.05 (0.02–0.13)Direct AIS construct subset (Sensitivity analyses)Direct AIS subset90.92 (0.86–0.96)0.89 (0.81–0.94)94.73 (31.09–288.67)8.24 (4.63–14.69)0.09 (0.04–0.17)Pooled sensitivities (Se) and specificities (Sp) with 95% CIs were obtained from random–effects bivariate/HSROC models. DOR and likelihood ratios (LR + and LR–) are derived from the bivariate summary point. Subgroup rows correspond to prespecified moderators (Posterior fossa Included, reference standard robustness, validation type) analyzed in separate bivariate fits. Abbreviations: Se, sensitivity; Sp, specificity; LR+, positive likelihood ratio; LR–, negative likelihood ratio; DOR, diagnostic odds ratio; CI, confidence interval Summary of pooled estimates Pooled sensitivities (Se) and specificities (Sp) with 95% CIs were obtained from random–effects bivariate/HSROC models. DOR and likelihood ratios (LR + and LR–) are derived from the bivariate summary point. Subgroup rows correspond to prespecified moderators (Posterior fossa Included, reference standard robustness, validation type) analyzed in separate bivariate fits. Abbreviations: Se, sensitivity; Sp, specificity; LR+, positive likelihood ratio; LR–, negative likelihood ratio; DOR, diagnostic odds ratio; CI, confidence interval HSROC Plot Subgroup analyses Posterior fossa: Including posterior fossa was associated with higher sensitivity and slightly lower specificity versus anterior-only cohorts; overall discrimination favoured inclusion. Posterior fossa: Including posterior fossa was associated with higher sensitivity and slightly lower specificity versus anterior-only cohorts; overall discrimination favoured inclusion. Reference-standard robustness: Pooled performance was similar under robust versus mixed/weaker standards; no consistent difference emerged. Reference-standard robustness: Pooled performance was similar under robust versus mixed/weaker standards; no consistent difference emerged. Validation type: External validation tended to show lower sensitivity with broadly comparable specificity relative to internal splits, consistent with expected generalisability effects. Validation type: External validation tended to show lower sensitivity with broadly comparable specificity relative to internal splits, consistent with expected generalisability effects. Meta-regression (univariable) Posterior fossa (excluded vs. included): Excluding posterior fossa was associated with a significant decrease in sensitivity, while specificity did not differ significantly (Table 4). Posterior fossa (excluded vs. included): Excluding posterior fossa was associated with a significant decrease in sensitivity, while specificity did not differ significantly (Table 4). Reference-standard robustness (mixed/weaker vs. robust): No significant associations were detected for either sensitivity or specificity. Reference-standard robustness (mixed/weaker vs. robust): No significant associations were detected for either sensitivity or specificity. Validation type (internal vs. external): Internal validation was associated with higher sensitivity; the specificity difference was not significant. Validation type (internal vs. external): Internal validation was associated with higher sensitivity; the specificity difference was not significant. Table 4Univariable meta-regressionModeratork_Setau2_SeI2%_SeSe_level0Se_level1Se_Δprob(1 − 0)Se_moderator_logitSe_interpretationk_Sptau2_SpI2%_SpSp_level0Sp_level1Sp_Δprob(1 − 0)Sp_moderator_logitSp_interpretationPosterior fossa included110.34864.5Included(0): 0.896 (95% CI 0.819–0.943)Excluded(1): 0.698 (95% CI 0.537–0.822)−0.198−1.316 (95% CI −2.264– −0.368); p = 0.007Excluded(1) shows lower sensitivity than Included(0) (Δ≈−0.198; 0.90→0.70); statistically significant (p = 0.007). Heterogeneity: I²=64.5% (substantial).110.6688.1Included(0): 0.861 (95% CI 0.753–0.926)Excluded(1): 0.924 (95% CI 0.835–0.966)0.0630.671 (95% CI −0.449– 1.792); p = 0.240Excluded(1) shows higher specificity than Included(0) (Δ≈+0.063; 0.86→0.92); not statistically significant (p = 0.240). Heterogeneity: I²=88.1% (high).Reference standard robustness131.00787.8Robust(0): 0.882 (95% CI 0.751–0.948)MixedWeaker(1): 0.855 (95% CI 0.702–0.936)−0.027−0.237 (95% CI −1.520– 1.045); p = 0.717MixedWeaker(1) shows lower sensitivity than Robust(0) (Δ≈−0.027; 0.88→0.85); not statistically significant (p = 0.717). Heterogeneity: I²=87.8% (high).130.77188.8Robust(0): 0.881 (95% CI 0.774–0.941)MixedWeaker(1): 0.904 (95% CI 0.818–0.952)0.0230.245 (95% CI −0.819– 1.309); p = 0.652MixedWeaker(1) shows higher specificity than Robust(0) (Δ≈+0.023; 0.88→0.90); not statistically significant (p = 0.652). Heterogeneity: I²=88.8% (high).Validation type120.79485.1External(0): 0.807 (95% CI 0.660–0.900)Internal(1): 0.940 (95% CI 0.851–0.978)0.1331.327 (95% CI 0.051–2.603); p = 0.041Internal(1) shows higher sensitivity than External(0) (Δ≈+0.133; 0.81→0.94); statistically significant (p = 0.041). Heterogeneity: I²=85.1% (high).120.8789.8External(0): 0.883 (95% CI 0.775–0.943)Internal(1): 0.909 (95% CI 0.807–0.960)0.0260.276 (95% CI −0.897– 1.449); p = 0.644Internal(1) shows higher specificity than External(0) (Δ≈+0.026; 0.88→0.91); not statistically significant (p = 0.644). Heterogeneity: I²=89.8% (high).Using logit-PLO REML models and restricting moderators to 0/1 (dropping “unclear”), exclusion of the posterior fossa (anterior-only; k=11) was associated with lower sensitivity (Δlogit −1.32, p=0.007; predicted 0.896→0.698) with no significant change in specificity (Δlogit +0.67, p=0.240; 0.861→0.924); heterogeneity was moderate for sensitivity (I²=64.5%) and high for specificity (I²=88.1%). Reference-standard robustness (robust vs mixed/weaker; k=13) showed no material effect on sensitivity (Δlogit −0.24, p=0.717; 0.882→0.855) or specificity (Δlogit +0.25, p=0.652; 0.881→0.904), with high heterogeneity (I²≈88–89%). Validation type (external vs internal; k=12) showed higher sensitivity with internal validation (Δlogit +1.33, p=0.041; 0.807→0.940) and no significant difference in specificity (Δlogit +0.28, p=0.644; 0.883→0.909); heterogeneity remained high (I²≈85–90%) Univariable meta-regression Using logit-PLO REML models and restricting moderators to 0/1 (dropping “unclear”), exclusion of the posterior fossa (anterior-only; k=11) was associated with lower sensitivity (Δlogit −1.32, p=0.007; predicted 0.896→0.698) with no significant change in specificity (Δlogit +0.67, p=0.240; 0.861→0.924); heterogeneity was moderate for sensitivity (I²=64.5%) and high for specificity (I²=88.1%). Reference-standard robustness (robust vs mixed/weaker; k=13) showed no material effect on sensitivity (Δlogit −0.24, p=0.717; 0.882→0.855) or specificity (Δlogit +0.25, p=0.652; 0.881→0.904), with high heterogeneity (I²≈88–89%). Validation type (external vs internal; k=12) showed higher sensitivity with internal validation (Δlogit +1.33, p=0.041; 0.807→0.940) and no significant difference in specificity (Δlogit +0.28, p=0.644; 0.883→0.909); heterogeneity remained high (I²≈85–90%) Additional analyses External validation only: Sensitivity was lower while specificity was broadly comparable to the primary model. External validation only: Sensitivity was lower while specificity was broadly comparable to the primary model. Robust reference standards only: Estimates closely matched the base analysis, indicating robustness to reference-standard choice. Robust reference standards only: Estimates closely matched the base analysis, indicating robustness to reference-standard choice. Posterior fossa included (excluding anterior-only) Sensitivity improved with a small trade-off in specificity, consistent with subgroup/meta-regression patterns. Posterior fossa included (excluding anterior-only) Sensitivity improved with a small trade-off in specificity, consistent with subgroup/meta-regression patterns. Direct AIS construct only (patient-level AIS detection; excluding CTA-defined LVO/HAS and ASPECTS-only tools): Performance was essentially unchanged from the base HSROC, with high sensitivity, high specificity, and strong overall discrimination. Direct AIS construct only (patient-level AIS detection; excluding CTA-defined LVO/HAS and ASPECTS-only tools): Performance was essentially unchanged from the base HSROC, with high sensitivity, high specificity, and strong overall discrimination. Discussion Principal findings Across included studies, DL algorithms applied to NCCT show high sensitivity and high specificity for AIS detection, with performance generally exceeding historical radiologist sensitivity on early CT signs and posterior fossa infarcts and approaching MRI benchmarks in in selected contexts [2, 4, 30]. External validation cohorts consistently showed lower sensitivity than internal tests, underscoring domain shift and generalization gaps that are well-documented in medical imaging AI [31]. Comparison with prior reviews Our results align with broader reviews of AI in stroke imaging that highlight promising accuracy but limited prospective and external validation [11]. Importantly, an NCCT-specific meta-analysis reached similar conclusions: strong internal validation performance contrasted by substantially attenuated sensitivity on external datasets and notable heterogeneity and bias risks [32]. Together, these syntheses and our patient-level meta-analysis converge on the same message: DL for NCCT-AIS is accurate in principle, but robust generalizability remains the rate-limiting step to clinical translation. Clinical Implications (rule-out vs. rule-in; workflow impact) Given NCCT’s speed and ubiquity, DL may be best positioned as a rule-in adjunct to highlight subtle early ischemic changes (and posterior fossa infarcts) that are often missed by humans, potentially expediting CTA/CTP/MRI and treatment decisions [2, 4, 30]. Caution is warranted for rule-out use, especially in settings with domain shift; thresholding should reflect local disease prevalence and downstream actions. To quantify net clinical value beyond AUC and accuracy, decision-curve analysis (DCA) should be routinely reported; this clarifies when an AI-assisted pathway yields higher net benefit than “scan all” or “scan none” strategies across clinically plausible thresholds [33, 34]. DL can also reshape workflows: (i) pre-read triage and worklist prioritization; (ii) real-time decision support for non-neuroradiologists in EDs without 24/7 subspecialty coverage; and (iii) standardized quantification of early ischemic change to support thrombolysis/thrombectomy selection. Sites adopting AI should pair deployment with local calibration and ongoing performance monitoring. Reporting gaps and study limitations Methodological and reporting quality varied. Frequent issues included unclear reference standards, leakage between development and test sets, minimal error analysis, and sparse reporting of model versioning, governance, and human–AI interaction. Adherence to CLAIM, CONSORT-AI/SPIRIT-AI, PRISMA-DTA for reviews, and QUADAS-2 for primary diagnostic studies would materially improve transparency and reproducibility [12–15]. Generalizability was the dominant limitation: performance declined on external cohorts, consistent with known site- and scanner-shift effects [31]. Equity risks are non-trivial: AI systems encode latent demographic signals (including race) even when these are not visible to experts, creating potential for disparate performance if training data are geographically or demographically narrow [35]. Many included studies were developed primarily on datasets from high-income, Western settings; broader international, multi-vendor datasets and subgroup reporting are needed. Strengths of this review We conducted a comprehensive, protocol-driven synthesis with patient-level meta-analysis where feasible, emphasized external validation and subgroup analyses, and interpreted accuracy alongside potential clinical utility (via DCA concepts) rather than accuracy metrics alone. We also mapped reporting to contemporary guidelines to facilitate reproducibility and appraisal [12–15]. Future Research Directions Prospective and pragmatic trials should evaluate AI-assisted NCCT pathways on patient-centered outcomes, designed and reported per CONSORT-AI/SPIRIT-AI and early-stage pilot work aligned with DECIDE-AI; future diagnostic-accuracy papers should follow STARD-AI [14, 15, 36, 37]. Explainability and failure analysis: standardized error taxonomies, case-level rationales, and model versioning. Clinical utility & economics: routine DCA (with uncertainty intervals) and cost-effectiveness analyses tied to local prevalence and downstream imaging/treatment cascades [33, 34]. Generalizability & equity: multi-region, multi-vendor datasets; pre-specified subgroup analyses; fairness audits given evidence that imaging AI can encode protected characteristics [35]. Deployment science: calibration drift monitoring, human–AI teaming studies, and learning-health-system feedback loops. Conclusion DL applied to NCCT shows high diagnostic accuracy for AIS and clear integration potential for triage and decision support. The core gaps are rigorous reporting, external validity, equity, and real-world impact evaluation. Addressing these with guideline-concordant prospective studies and utility/economic analyses should enable safe, effective adoption. Supplementary Information Supplementary Material 1. Acknowledgements The authors would like to thank the librarians and information specialists who provided assistance during the systematic search. Institutional review board statement Not applicable. Informed consent statement Not applicable. Conflicts of interest The authors declare no conflict of interest. Authors’ contributions Conceptualization, Kalab Yigermal Gete and Asnakew Achaw Ayele; Methodology, Kalab Yigermal Gete and Asnakew Achaw Ayele; Software, Kalab Yigermal Gete; Validation, Kalab Yigermal Gete and Asnakew Achaw Ayele; Formal analysis, Kalab Yigermal Gete; Investigation, Kalab Yigermal Gete and Asnakew Achaw Ayele; Resources, Kalab Yigermal Gete; Data curation, Kalab Yigermal Gete and Asnakew Achaw Ayele; Writing—original draft, Kalab Yigermal Gete; Writing—review & editing, Kalab Yigermal Gete and Asnakew Achaw Ayele; Visualization, Kalab Yigermal Gete; Supervision, Asnakew Achaw Ayele; Project administration, Kalab Yigermal Gete; Funding acquisition, Not applicable. Study selection and data extraction were performed independently and in duplicate by Kalab Yigermal Gete and Asnakew Achaw Ayele; disagreements were resolved by consensus. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability All data generated or analyzed during this study are included in this published article and its supplementary materials. Additional data extraction sheets and meta-analysis code are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement All data generated or analyzed during this study are included in this published article and its supplementary materials. Additional data extraction sheets and meta-analysis code are available from the corresponding author upon reasonable request.