Performance of an artificial intelligence tool for multi-step acute stroke imaging: A multicenter diagnostic study

Abstract

Objective

Several artificial intelligence (AI) tools have been developed to assist in the stroke imaging workflow, which remains a major disease of the 21st century. This study evaluated the combined performance of an FDA-cleared and CE-marked AI-based device with three modules designed to detect intracerebral hemorrhage (ICH), identify large vessel occlusion (LVO), and calculate Alberta Stroke Program Early CT Scores (ASPECTS).

Materials & methods

Non-contrast CT (NCCT) and/or computed tomography angiography (CTA) for suspicion of stroke acquired at La Timone and Nord University hospitals (Marseille, France) between March 2019 and March 2020 were retrospectively collected. The AI tool, CINA-HEAD (Avicenna.AI), processed the data to flag ICH, LVO, and calculate ASPECTS. The results were compared to ground truth evaluations by four expert neuroradiologists to compute diagnostic performances.

Results

A total of 373 NCCT and 331 CTA from 405 patients (mean age 64.9 ± 18.9 SD, 52.6 % female) were included. The AI tool achieved an accuracy of 94.6 % [95 % CI: 91.8 %-96.7 %] for ICH detection on NCCT and of 86.4 % [95 % CI: 82.2 %-89.9 %] for LVO identification on CTA. The region-based ASPECTS analysis yielded an accuracy of 88.6 % [95 % CI: 87.8 %-89.3 %] and the dichotomized ASPECTS classification (ASPECTS ≥ 6) achieved 80.4 % accuracy.

Conclusion

This study demonstrates the reliable, stepwise performance of an AI-based stroke imaging tool across the diagnostic cascade of ICH and LVO detection and ASPECTS scoring. Such robust multi-stage evaluation supports its potential for streamlining acute stroke triage and decision-making.

Highlights

• •The AI tool demonstrated high accuracy in ASPECT scoring, ICH and LVO detection.
• •Performance was consistent across ICH subtypes, ensuring accurate rule-in rule-out.
• •Moderate sensitivity, high specificity ensured LVO rule-in, reducing false positives for ASPECTS.
• •The AI tool may align with stroke imaging workflow enhancing clinical decision-making.

1. Introduction

Despite advances in prevention and treatment, stroke remains a major health issue, affecting 15 million people annually and causing 5 million deaths [1]. In the past 20 years, the global burden of stroke continued to rise, with a 70 % increase in incidence and a 102 % increase in prevalence over the past two decades [2], [3].

Clinically, strokes are broadly categorized into ischemic and hemorrhagic subtypes. Ischemic stroke (IS), caused by arterial occlusion, accounts for approximately 80 % of cases, while hemorrhagic stroke, resulting from vascular rupture and causing spontaneous intracranial hemorrhage (ICH), comprises the remaining 20 % [4], [5], [6]. This distinction is crucial, as it dictates different imaging approaches and treatment pathways in the acute setting.

In stroke treatment, the principle "time is brain" underscores the importance of rapid diagnosis and prompt intervention. For every minute a stroke remains untreated, nearly 2 million neurons and 14 billion synapses are lost, equating to 3.6 years of brain aging for every untreated hour [7], [8]. Hence, rapid neuroimaging plays a central role in triaging patients for adequate treatment such as thrombolysis or thrombectomy.

Many emergency departments use predefined stroke protocols to optimize patient outcomes. Upon initial patient evaluation, a non-contrast CT (NCCT) is conducted to rule out ICH, followed by a computed tomography angiography (CTA) to detect vessel occlusions [9], [10], [11]. If a large vessel occlusion (LVO) is confirmed in the middle cerebral artery (MCA) and/or the internal carotid artery (ICA), the Alberta Stroke Program Early CT Score (ASPECTS) is used to quantify early ischemic changes (EIC) on NCCT. The ASPECTS method deducts one point from the initial score of 10 for every region of the MCA vascular territory (M1, M2, M3, M4, M5, M6, internal capsule - IC, caudate - C, lentiform -L and insula - I) exhibiting EIC. According to American and European guidelines, patients with ASPECTS ≥ 6 are strong candidates for endovascular treatment due to significantly improved outcomes [12], [13]. Furthermore, depending on departmental resources, additional imaging modalities such as CT perfusion or magnetic resonance imaging may be employed for further assessments. Thus, imaging is not interpreted in isolation; each imaging step informs the next action.

In parallel with improvements in imaging protocols, Artificial Intelligence (AI) applications have emerged as a meaningful resource to support clinicians across the stroke workflow. They are a valuable alternative to improve inter-rater agreement, which is often only moderate in ASPECTS assessment and influenced by factors such as reader experience, image quality, and CT reconstruction techniques [14], [15], [16]. Moreover, AI tools can help reduce missed diagnoses, reported in up to 4.3 % of ICH cases and 10 % of LVO cases [17], [18], and decrease interpretation time, thereby supporting faster and more consistent clinical decision-making [19], [20], [21], [22]. Multiple retrospective studies, both single-center and multicenter, have evaluated the diagnostic performance of individual AI tools. Reported accuracies range from 93 % to 96 % for ICH detection [23], [24], [25], [26], [27], [28], [29], [30], [31], [32] and from 81 % to 95 % for LVO detection [33], [34], [35], [36], [37], [38], [39], [40], [41], [42]. Regarding automated ASPECTS assessment, studies have demonstrated moderate-to-good performance, with accuracies ranging from 61 % to 87 % [19], [43], [44], [45], [46], [47], [48].

While individual performance metrics are encouraging, they do not reflect the clinical context in which the tools are used. The integrated performance of a multi-task AI system across the full imaging workflow (ICH detection, LVO identification, ASPECTS computation) has not been evaluated. Indeed, since these AI tasks are used in a sequential decision-making process, evaluating their integrated performance is crucial, as the accuracy of each step directly influences the next and errors may cascade, compromising subsequent decisions. For example, accurate ICH triage is critical at the outset: ruling out ICH is a prerequisite for proceeding to CTA and evaluating for LVO, as ASPECTS is not applied in cases with hemorrhage. Similarly, reliable LVO triage is necessary to rule in LVO before ASPECTS scoring becomes clinically relevant. This sequential dependency highlights the importance of deploying AI-based tools within the clinical routine in a way that mirrors real decision-making processes, not only to optimize individual task accuracy, but to support the entire diagnostic pathway.

To address this gap, this study aims to present the performance analysis of an FDA-cleared and CE-marked AI tool integrated into the stroke imaging workflow of two large comprehensive stroke centers. The system includes three applications designed to detect ICH, identify LVO, and compute ASPECT scores. We evaluate the tool’s integrated diagnostic performance compared to neuroradiologist interpretation, with the aim of understanding its potential to support clinical decision-making in real-world settings.

2. Materials and methods

2.1. Data collection

Data were retrospectively and consecutively collected from La Timone and Nord University Hospitals (Marseille, France) between March 2019 and March 2020. The study included all patients who underwent NCCT and/or CTA scans following the stroke imaging workflow (Fig. 1.a). Experts who established the ground truth excluded cases according to the following criteria: age under 18 years, post-neurosurgery patients, significant image artifacts (uninterpretable images), supra-aortic trunk not visible in CTA cases limiting LVO analysis, and bilateral or chronic stroke for ASPECTS analysis. Finally, cases that could not be processed by the AI application due to non-compliance with required acquisition parameters, such as strict axial acquisition, absence of interslice gaps, field of view above 170 mm (for ASPECTS) and use of a soft tissue reconstruction kernel, were excluded from the initial assessment.

Fig. 1.

Stroke imaging workflow (a) and flow diagram of participants (b) for ICH, LVO and ASPECTS.

CT scans were acquired on two General Electric Revolution EVO (128-slice) CT scanners of the same model. The X-ray tube voltage was 100 or 120 kVp, depending on the patient’s physical build. Contrast agents used for CTA were Omnipaque 350 (GE Healthcare) or Iomeron 350 (Bracco) at standard injection rate of 4 mL/s and total volume of 60 mL, or Visipaque 320 (GE Healthcare) at standard injection rate of 5 mL/s and volume up to 70 mL. A slice thickness of 1.0 mm was used for the AI-based application. All CTA studies were acquired using a single-phase protocol.

2.2. Ground truth

Retrospective analysis of the CT scans was performed first by consensus among three senior neuroradiologists, T.A, A.R. and A.E, with 3 years, 17 years and 7 years of experience in neuroradiology, respectively. All first evaluations were conducted jointly, with the experts reviewing each case simultaneously and agreeing on the findings through discussion. They evaluated the presence of ICH on NCCT and categorized their subtypes (intraparenchymal - IPH, intraventricular - IVH, subarachnoid - SAH, subdural and/or epidural - SDH/EDH). Moreover, the presence and location of LVO within proximal ICA, MCA-M1 or distal MCA-M2 segments was evaluated on CTA. In cases of ischemic stroke, the ASPECT score was determined by visual assessment of EIC in each of the ASPECTS regions on NCCT images. When available, the experts used the CTA to confirm the infarct location. Cases with posterior circulation stroke without any occlusion in the MCA or ICA segments were assessed as negative for ASPECTS (ASPECTS = 10) and LVO, because they are not targeted by the application.

Then, the same cases were processed by the AI tool CINA-HEAD (Avicenna.AI, La Ciotat, France). An expert neuroradiologist (G.B) with 15 years experience in neuroradiology reviewed all the cases in order to establish the final ground truth (GT). Discrepancies between the consensus of the first three neuroradiologists and the AI tool were resolved by majority agreement. Patient clinical history was accessible to the evaluators when available.

2.3. AI-based application

All imaging data were processed using CINA-HEAD (Avicenna.AI, La Ciotat, France), an FDA-cleared and CE-marked AI-based tool. This software comprises three algorithms designed to flag suspected ICH cases on NCCT (CINA-ICH v1.0), suspected LVO cases on CTA (CINA-LVO v1.0), and calculate the ASPECT Score on NCCT (CINA-ASPECTS v1.4.1).

If a case is flagged with suspicion of ICH, the algorithm generates an automated notification accompanied by a secondary capture image that outlines the hemorrhage using a contour. The tool is designed to detect all types of acute intracranial hemorrhages. For CINA-LVO, a similar notification is sent when a LVO is identified, along with a secondary capture highlighting the suspected occlusion using a bounding box. This tool was designed to detect occlusions in specific segments of the anterior circulation: the M1 segment of the middle cerebral artery (MCA), the proximal MCA-M2 segment, and the distal internal carotid artery (ICA). The goal of both CINA-ICH and CINA-LVO is to support rapid triage and optimize workflow prioritization in acute stroke care. Regarding CINA-ASPECTS, the algorithm produces a map identifying the 20 ASPECTS regions, with each region outlined in green (indicating no EIC) or red (indicating EIC). A transparent red overlay also highlights affected areas on the NCCT series. In addition, the application provides a table summarizing the average Hounsfield Unit values for each region, along with the computed ASPECT score and the affected hemisphere. Variations of the ASPECT scoring system such as posterior circulation ASPECTS (pc-ASPECTS) are not assessed by this tool.

All three algorithms are based on convolutional neural networks (CNNs) using hybrid 3D/2D Unet topologies [49], [50]. The ICH and LVO classifiers were trained on 10,345 and 20,161 cases, respectively, sourced from several U.S. clinical centers. The data were adequately distributed in terms of scanner vendors (Canon Medical Systems Corporation, GE Healthcare, Philips Healthcare, Siemens Healthineers), patient age, slice thickness, and kVp. During their validation phases, both classifiers achieved accuracies exceeding 95.6 % [30]. The ASPECTS algorithm was trained on 2097 NCCT cases also adequately distributed, achieving a per-region accuracy of 87.0 % during the validation phase [51].

2.4. Statistical analysis

Firstly, the inter-rater agreement between the initial consensus of the three neuroradiologists and the final expert neuroradiologist was evaluated using Cohen’s kappa [52]. Then, to reflect the real-world clinical workflow, the AI tool's performance was evaluated in a cascade manner, following the standard imaging sequence. Cases were initially analyzed by CINA-ICH and based on the result, they were passed to CINA-LVO for occlusion analysis. If an LVO was identified, CINA-ASPECTS then evaluated early ischemic changes.

The algorithm results for ICH and LVO were compared to the GT, and the accuracy (overall agreement), sensitivity, specificity and the area under the receiver operating curve (ROC AUC) were calculated on a per-case basis. Additionally, the performance per ICH type and per LVO location were evaluated.

For ASPECTS analysis, region-based performance was assessed against the GT for all regions and for each individual ASPECTS region. The accuracy was computed for cortical (M1-M6) and deep (I, L, C, IC) regions. Furthermore, a score-based analysis was computed and the intraclass correlation coefficient (ICC) between the AI algorithm was evaluated. Finally, a dichotomized analysis using a cutoff point of ASPECTS ≥ 6 (endovascular cut-off point) was also performed and the percentage of cases correctly classified as ASPECTS ≥ 6 by the AI was compared to the GT.

All 95 % confidence intervals (95 % CI) were calculated using the exact binomial distribution (Clopper Pearson) [53]. All the statistical analyses were performed using MedCalc Statistical Software (v20.015, MedCalc Software Ltd).

3. Results

A total of 405 patients (mean age 64.9 ± 18.9 [SD], 52.6 % female) met the initial inclusion criteria. Regarding ICH evaluation, 16 cases were not processed by the algorithm and another 16 cases were excluded by the experts who established the GT due to significant artifacts rendering images uninterpretable. This resulted in 373 NCCT cases included for analysis. For LVO analysis, 10 cases were not processed by the algorithm, 60 cases lacked visible supra-aortic trunk limiting LVO analysis, and 4 cases had significant artifacts, yielding 331 cases included. Finally, for ASPECTS analysis, 34 cases were not processed by the algorithm, 2 were chronic strokes, and 17 involved bilateral strokes, yielding 352 NCCT cases. Among these, 282 cases had corresponding CTA images, which were used to confirm the side of infarction (left or right). Fig. 1.b outlines the workflow for case inclusion and exclusion. The specific reasons for AI processing failures are detailed in Table S1 of the supplementary material.

Regarding the GT, discrepancies between the initial consensus of the three neuroradiologists and the final expert neuroradiologist were observed in 6 out of 373 cases (1.6 %) for ICH, resulting in a Cohen’s kappa of 0.964 (95 % CI: 0.929–0.992). For LVO, 10 out of 331 cases (3.0 %) were in disagreement, yielding a kappa of 0.923 (95 % CI: 0.875–0.970). For ASPECTS, 260 out of 3520 evaluated regions (7.4 %) showed disagreement, with a corresponding kappa of 0.662 (95 % CI: 0.621–0.697). After resolving discrepancies and establishing the final GT, 28.6 % of NCCTs were positive for ICH, 25.1 % of CTAs were positive for LVO, and the mean ASPECTS across 352 NCCT cases was 9.3 ± 2.6 [SD].

3.1. Intracranial hemorrhage

For ICH, the analysis of 373 NCCT cases yielded an overall accuracy of the algorithm against the ground truth of 94.6 % [95 % CI: 91.8 % - 96.7 %], a sensitivity of 86.0 % [95 % CI: 77.9–91.9 %], a specificity of 98.1 % [95 % CI: 95.7–99.4 %] and a ROC AUC of 0.921 [95 % CI: 0.888–0.946] (Fig. 2). There were 5 false positives attributed to hyperdense venous sinuses (n = 2), normal cortex near ischemic zones (n = 2), and significant motion artifacts (n = 1). Among 15 false negatives, there were 5 missed cases with SAH, 4 missed cases with SDH, 2 missed cases with IPH, and 4 with mixed presentations. These mixed presentations included 1 missed case with IPH and IVH, 1 missed case with IPH and EDH, 1 missed case with SDH and SAH, and 1 missed case with IPH and SAH. Accuracy across ICH subtypes exceeded 94.6 % (Table 1), with SAH presenting the lowest value (94.6 % [95 % CI: 91.8–96.7 %]) and IPH the highest value (97.3 % [95 % CI: 95.1–98.7 %]). Examples of outputs including a true positive, false positive, and false negative result are shown in Fig. 3(a–c).

Fig. 2.

Area under the receiver operating curve (AUC) for ICH (a), LVO (b) and ASPECTS.

Table 1.

Algorithm’s accuracy for overall ICH and ICH subtypes.

ICH evaluation	Number of positive cases (%)	Accuracy [95 % CI]	Sensitivity [95 % CI]	Specificity [95 % CI]
All types	107 (28.7 %)	94.6 % [91.8–96.7 %]	86.0 % [77.9–91.9 %]	98.1 % [95.7–99.4 %]
IPH	53 (14.2 %)	97.3 % [95.1–98.7 %]	83.0 % [70.2–91.9 %]	99.7 % [98.3–99.9 %]
IVH	38 (10.2 %)	96.5 % [94.1–98.1 %]	65.8 % [48.7–80.4 %]	100 % [98.9–100 %]
SAH	51 (13.7 %)	94.6 % [91.8–96.7 %]	78.4 % [64.7–88.7 %]	97.2 % [94.8–98.7 %]
SDH/EDH	43 (11.5 %)	95.4 % [92.8–97.3 %]	76.7 % [61.4–88.2 %]	97.9 % [95.7–99.1 %]

*Please note that one case can have more than one ICH subtype.

Fig. 3.

ICH detection examples include a true positive SDH (a), a false positive caused by a hyperdense venous sinus (b) and a false negative corresponding to a missed subtle SDH (c). LVO detection examples include a true positive M1 occlusion (d), a false positive due to a stenosis (e) and a false negative corresponding to a missed M2 proximal occlusion (f).

3.2. Large vessel occlusion

For LVO detection, the overall accuracy was 86.4 % [95 % CI: 82.2–89.9 %], sensitivity was 55.4 % [95 % CI: 44.1–66.3 %], specificity was 96.8 % [95 % CI: 93.7–98.6 %] and the ROC AUC (Fig. 2) was 0.761 (95 % Ci: 0.711–0.806]. There were 37 false negative cases with varying occlusion locations: LVO located in the M2 segment (n = 21), ICA segment (n = 13), M1 segment (n = 1), M1 and M2 segments (n = 1), and ICA, M1 and M2 segments (n = 1). False positives were observed in 8 cases, attributed to stenosis (n = 4), motion artifacts (n = 2), glioma (n = 1), and neurosurgical clips (n = 1). Segment-specific accuracies for ICA, M1 and M2 segments were 95.5 %, 94.9 %, 85.0 %, respectively (Table 2). Examples of LVO outputs are presented in Fig. 3(e-f).

Table 2.

Algorithm’s accuracy for overall LVO and each LVO segment.

LVO evaluation	Number of positive cases (%)	Accuracy [95 % CI]	Sensitivity [95 % CI]	Specificity [95 % CI]
All segments	83 (25.1 %)	86.4 % [92.2–89.9 %]	55.4 % [44.1–66.3 %]	96.8 % [93.7–98.6 %]
Proximal ICA	30 (9.1 %)	95.5 % [90.5–96.0 %]	53.3 % [34.3–71.7 %]	99.7 % [98.2–99.9 %]
MCA-M1	41 (12.4 %)	94.9 % [91.9–97.0 %]	75.6 % [59.7–87.6 %]	97.6 % [95.1–99.0 %]
Distal MCA-M2	65 (19.6 %)	85.0 % [80.6–88.6 %]	53.9 % [41.0–66.3 %]	92.5 % [88.6–95.4 %]

*Please note that one case can present more than one LVO segment.

3.3. Aspects

Regarding ASPECTS, the region-based accuracy of the algorithm for all 352 NCCT cases was 88.6 % [95 % CI: 87.8–89.3 %], sensitivity was 68.0 % [95 % CI: 61.5–73.9 %], specificity was 89.3 % [88.5–90.0 %] and the ROC AUC was 0.786 [95 % CI: 0.776–0.796]. Cortical and deep regions obtained accuracies of 89.6 % [95 % CI: 88.6–90.5 %] and 87.1 % [95 % CI: 85.8–88.3 %], respectively. The lowest accuracy was in the lentiform (80.3 %) and the highest was in the caudate (91.8 %). All values are reported in Table 3 and an example of ASPECTS output is shown in Fig. 4.

Table 3.

Algorithm’s accuracy across all ASPECTS regions and for each ASPECTS region.

ASPECTS Region	Number of regions with EIC (%)	Accuracy [95 % CI]	Sensitivity [95 % CI]	Specificity [95 % CI]
All Regions	231 (3.3 %)	88.6 % [87.8–89.3 %]	68.0 % [61.5–73.9 %]	89.3 % [88.5–90.0 %]
M1	15 (2.1 %)	90.3 % [87.9–92.4 %]	66.7 % [38.4–88.2 %]	90.9 % [88.5–92.9 %]
M2	25 (3.6 %)	86.7 % [83.9–89.1 %]	84.0 % [63.9–95.5 %]	86.8 % [84.0–89.2 %]
M3	20 (2.8 %)	91.2 % [88.9–93.2 %]	65.9 % [40.8–84.6 %]	92.0 % [89.7–93.9 %]
M4	16 (2.3 %)	91.5 % [89.2 % - 93.4 %]	43.8 % [19.8–70.1 %]	92.6 % [90.4–94.4 %]
M5	26 (3.7 %)	89.4 % [86.8 % - 91.5 %]	53.9 % [33.4–73.4 %]	90.7 % [88.3–92.8 %]
M6	23 (3.3 %)	88.4 % [85.8–90.6 %]	69.6 % [47.1–86.8 %]	89.0 % [86.4–91.2 %]
I	33 (4.7 %)	88.8 % [86.2–91.0 %]	75.8 % [57.7–88.9 %]	89.4 % [86.8–91.6 %]
IC	18 (2.6 %)	87.6 % [85.0–90.0 %]	50.0 % [26.0–74.0 %]	88.6 % [86.0–90.9 %]
L	30 (4.3 %)	80.3 % [77.1–83.1 %]	83.3 % [65.3–94.4 %]	80.1 % [76.9–83.0 %]
C	25 (3.6 %)	91.8 % [89.5–93.7 %]	68.0 % [46.5–85.1 %]	92.6 % [90.4–94.5 %]

Fig. 4.

Example of an ASPECTS case. The AI-based application identified EIC in M1,M2, M3, M4, M5 and M6 (ASPECTS = 4) whereas the GT identified EIC in M2, M3, M4, M5 and M6 (ASPECTS = 5). (a), (c), (d) and (e) show the outputs of the AI-based algorithm, and (b) and (d) the raw images.

A score-based analysis of all 352 ASPECT scores found that the ICC between the AI algorithm and the ground truth was 0.618, indicating good agreement. Additionally, scores were dichotomized at a threshold of ASPECTS ≥ 6. The algorithm was capable of correctly classifying 273/339 (80.5 %) cases as ASPECTS ≥ 6 and 10/13 (76.9 %) as ASPECTS < 6, yielding a dichotomized accuracy of 80.4 %.

4. Discussion

Given that stroke remains a major health issue of the 21st century, many emergency departments implement predefined stroke imaging protocols to ensure optimal patient outcomes. This study evaluated the performance of an FDA-cleared and CE-marked AI tool to assist clinicians across every step of the stroke imaging workflow, focusing on ICH detection, LVO identification and ASPECT Score calculation. Using 373 NCCT and 331 CTA scans from 405 patients, the AI tool demonstrated high accuracy within the stroke protocol, highlighting its potential to enhance workflow and eventually improve patient care.

ICH detection is typically the first diagnostic step in the stroke workflow, hence, an accurate rule-in and rule-out technique is crucial for guiding subsequent diagnostic steps effectively. CINA-HEAD tool obtained an accuracy of 94.64 % on 373 NCCT scans, with a moderately balanced sensitivity and specificity. This high performance remained consistent across all ICH subtypes (IPH, IVH, SAH and SDH/EDH). A relatively lower sensitivity was noted for IVH, though this was balanced by a very high specificity. Previous studies have evaluated other commercially available AI tools for ICH detection on NCCT and reported accuracy levels between 93 % and 95.5 % [23], [24], [25], [26], [27]. Regarding the current tool evaluated in this study, CINA-ICH, the published accuracy in the literature ranged from 92.5 % to 95.6 % [30], [31], [32]. These prior findings align closely with the current results, supporting the reliable performance of these types of AI-based screening tools for ICH detection.

LVO detection on CTA is usually the next step in the process to confirm the presence of IS, which accounts for nearly 80 % of all strokes [6]. High specificity is particularly important at this stage to accurately rule in and confirm the diagnosis of LVO and facilitate progression to subsequent analyses, such as ASPECTS scoring or advanced imaging techniques. The current findings demonstrated high specificity with moderate sensitivity, yielding a good accuracy of 86.4 % for CINA-LVO, and aligning with prior studies that report specificity values exceeding 88 % and sensitivities ranging from 52 % to 80 % [33], [34], [35], [36], [37], [38], [39], [40], [41], [42]. Low sensitivity in LVO detection is often attributed to challenges in detecting LVOs in MCA-M2 segments, a limitation also observed in this study. While achieving both high sensitivity and specificity is ideal, prioritizing high specificity is crucial when ASPECTS and other ischemic stroke analyses are conducted post-LVO confirmation. This approach ensures accurate rule-in, thus reducing the burden of excessive false positives on radiologists and increasing the likelihood of identifying true positives.

Unlike ICH and LVO triage applications, ASPECTS provides a quantitative assessment of the extent of EIC in the MCA vascular territory to guide appropriate endovascular treatment decisions. In this study, the CINA-ASPECTS tool achieved a region-based accuracy of 88.6 %, 89.6 % for cortical regions and 87.1 % for deep; and a score-based dichotomized accuracy (threshold ≥ 6) of 80.4 %, demonstrating sufficient reliability for patient selection. This high accuracy likely reflects the prior robust rule-in and rule-out steps for ICH and LVO, which ensure a refined input for ASPECTS evaluation. While previous studies have reported AI-assisted ASPECTS region-based accuracies ranging from 61 % to 87 % [19], [43], [44], [45], [46], [47], [48], these evaluations were limited to standalone AI performance. To the best of our knowledge, this is the first study assessing the integrated performance of an AI tool across the entire stroke workflow, from ICH and LVO screening to ASPECTS computation. The findings emphasize the positive impact of comprehensive and sequential rule-in and rule-out strategies in stroke diagnosis.

Indeed, imaging workflows are critical in the early treatment of stroke for determining the type, location, and severity of the stroke, enabling clinicians to select the most appropriate intervention. Integrated workflows supported by automated decision support tools have the potential to reduce delays in diagnosis and treatment, thus improving patient outcomes. For ICH, accurate detection facilitates prompt blood pressure management, enabling well-managed medical treatment or surgery for evacuation [54], [55]. Regarding IS, every minute of delay or misdiagnosis results in a loss of nearly 2 million neurons, 14 billion synapses, and 12 kilometers of myelinated fibers [6], [8]. Hence, the rapid identification of LVO, followed by timely ASPECTS scoring, can significantly reduce treatment delays, shorten hospital stays, minimize rehabilitation needs, enhance quality of life, and lower overall healthcare costs.

Recent studies show that advanced medical imaging, beyond assessing treatment efficiency, provides critical vascular and physiological information that strongly influences endovascular reperfusion outcomes, emphasizing its central role in patient care [6]. Indeed, advances in AI-based imaging can provide prompt and accurate data for ICH and LVO detection, which may enhance patient triage and improve outcomes. Previous studies on AI-based ICH detection and worklist prioritization have demonstrated enhanced efficiency, reduced turnaround times and shortened hospital stays, all of which facilitate better prognosis [23], [54], [56]. Similarly, automated LVO detection and ASPECTS computation have shown a notable impact in hub and spokes networks. The improvement was basically driven by the reduction of unnecessary transfers, door-in to door-out time, door-to-puncture time, inter-observer discrepancies, and time to initiation of mechanical thrombectomy [20], [21], [57], [58]. Future studies should prospectively evaluate the implementation of all these tools together, as their coordinated integration is expected to optimize the door-in-door-out workflow as a whole, enhancing time efficiency and ensuring improved patient outcomes.

This study has several limitations. Its retrospective design may introduce bias and limits assessment of direct clinical impact, such as time savings or outcomes. Additionally, the AI tool does not assess distal occlusions (e.g., distal ICA, ACA, or posterior circulation), and it doesn’t evaluate ASPECTS against final infarct volume, LVO against DSA, or ICH against hematoma expansion. These limitations reflect the AI system’s intended role as a triage and prioritization aid rather than a comprehensive diagnostic tool. While this relatively large, multicenter cohort offers some insight into generalizability, prospective studies with larger and more diverse populations are needed to confirm broader applicability. Finally, an additional limitation is the exclusion of scans that could not be processed due to incompatibility with the AI tool’s acquisition protocol. Although these cases represent a small proportion of the dataset, they underscore the importance of developing more flexible and inclusive AI protocols for broader clinical applicability.

In summary, this study demonstrates the potential of an FDA-cleared and CE-marked AI tool to enhance stroke imaging workflows by accurately integrating ICH detection, LVO identification, and ASPECT Score computation. The tool showed high accuracy, confirming its utility in streamlining decision-making and improving patient care. Future prospective studies should explore broader applications and imaging techniques such as CTP or DWI, in order to evaluate the impact in diverse clinical settings.

CRediT authorship contribution statement

Anthony Reyre: Writing – review & editing, Validation, Supervision, Formal analysis, Data curation. Gilles Brun: Writing – review & editing, Validation, Supervision, Conceptualization. Ahmed-Ali El Ahmadi: Writing – original draft, Validation, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Thibault Agripnidis: Writing – original draft, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Angela Ayobi: Writing – original draft, Visualization, Validation, Software, Project administration, Methodology, Formal analysis. Sarah Quenet: Writing – review & editing, Validation, Software, Resources, Project administration, Conceptualization. Yasmina Chaibi: Writing – review & editing, Validation, Software, Resources, Conceptualization. Christophe Avare: Writing – review & editing, Validation, Software, Resources, Conceptualization. Alexis Jacquier: Writing – review & editing, Validation, Supervision, Conceptualization. Nadine Girard: Writing – review & editing, Validation, Supervision, Conceptualization. Jean-François Hak: Writing – review & editing, Validation, Supervision, Conceptualization.

Ethical statement

We certify that all procedures were performed in compliance with relevant laws and institutional guidelines. Institutional Review Board approval was obtained in March 2021 (number PADS19–309). Written informed consent was waived by the Institutional Review Board.

Funding

This work was supported by Avicenna.AI (La Ciotat, France).

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Angela Ayobi, Sarah Quenet, Yasmina Chaibi and Christophe Avare report a relationship with Avicenna.AI that includes: employment. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

Supplementary material