Harston, George

Introduction: AI-driven software has been recommended as a means of triaging suspected large vessel occlusions (LVOs) on CT angiography (CTA), and different solutions are now commercially available. This study compared the performance of two software packages, Brainomix 360 Triage LVO and Rapid LVO, in prospective, real-world use. Methods: Consecutive patients scanned for suspected stroke at 3 comprehensive stroke centres from May to July 2024 (3 months) were included in the study. CTA, where available, was processed in real time by both Rapid and Brainomix software. Ground truth on vessel occlusion was ascertained using clinical review of electronic medical records. Occlusion location was coded as intracranial internal carotid artery terminus (ICA-T), middle cerebral artery (MCA) M1 or MCA M2. If more than one occlusion was observed, the most proximal of these was taken as the primary vessel affected. Results: There were 561 suspected stroke patients scanned during the study period across the three sites. Of these, CTA was sent for processing by both Rapid LVO and Brainomix 360 Triage LVO for 359 cases. Cases were excluded if no contrast was detected (n=17), leaving 342 cases (Table 1). Of these, proximal LVO (ICA-T or MCA M1) was confirmed in 14 cases and MCA M2 occlusions were present in 6 cases. Table 2 shows the LVO detection results for all cases from the two software packages. For both devices, sensitivity was higher when considering only proximal occlusions (ICA-T or M1). Performance was similar overall, but there was a trend towards differing sensitivity for more distal occlusions. As a composite measure, area under the curve (AUC) was higher for Brainomix than Rapid (Figure 1) but this difference was not significant. The positive predictive value (PPV) was low for both software packages due to the low incidence of LVO in the patient population. Conclusions: This study demonstrated consistent performance of two AI software devices, Brainomix 360 Triage LVO and Rapid LVO, for detection of LVO on CTA imaging, in prospective, real-world use. The low number of positive cases in the dataset (N=21) is a limitation of the study. There was a trend towards different performance for more distal vessel occlusion locations. Overall, the sensitivity observed in this dataset was in line with reports from previous retrospective studies for both software packages and underlines the importance of prospective evaluation to understand real-world performance.

Introduction Software packages for quantification of CT perfusion parameters allow quick visualization of perfusion deficits to support clinical decision making in acute ischemic stroke (AIS). Several packages are now commercially available for automated quantification of ischemic core, penumbra volumes, and mismatch ratios to indicate eligibility for reperfusion therapies. This study explored the agreement between two software packages, Brainomix 360 e‐CTP and Rapid CTP, in prospective, real‐world use. Methods Consecutive patients scanned for suspected stroke at three comprehensive stroke centers over 90 days from May to July 2024 were included in the study. Eligibility for the current analysis was limited to discharge diagnosis of AIS and CTP imaging processed by both Brainomix and Rapid software. Software processing was conducted in real time after image acquisition. The presence of large vessel occlusion (LVO) was determined from review of electronic medical records. CTP core (CBF<30%) and penumbra (Tmax>6s) volumes from Brainomix and Rapid were recorded for each case. Mismatch ratios were calculated (penumbra÷core volume); positive mismatch was defined according to DEFUSE‐3 criteria (core<70ml, penumbra>15ml, mismatch ratio > 1.8). The correlations and linear regressions between Brainomix and Rapid were computed for core, penumbra and mismatch ratios, and their concordance for mismatch positivity was assessed. Results 137 patients were scanned and received a final diagnosis of AIS. Of these, only 38 had imaging processed by both Brainomix and Rapid. Median age was 69 years (interquartile range: 57‐77.5y); 16 (42%) were female; 13 (34%) had LVO. Linear models between Brainomix and Rapid's outputs for CBF and Tmax volumes are shown in Figure 1a and 1b. The correlation between software outputs for CBF volume was r=0.70 (p<.001), and for Tmax volume was r=0.72 (p<.001). Mismatch ratios were calculated for the 13 cases with LVO (Figure 1c). The linear model between mismatch ratios was not significant, with four Rapid outputs showing infinite mismatch ratios (due to having CBF volumes of 0ml). Concordance between software for mismatch positivity was good; 8/13 cases showed mismatch on both softwares; 1 case showed mismatch on Brainomix only; 1 for Rapid only; and 3 cases showed mismatch on neither software. Conclusion This prospective study, despite limited sample size (N=38), showed good correlations between Brainomix 360 e‐CTP and Rapid CTP for calculation of both core and penumbra volumes. Although mismatch ratios showed less agreement, the categorization of mismatch positivity showed good concordance between software platforms. image

Background: Identifying optimal candidates for endovascular treatment (EVT) remains an ongoing priority to maximize treatment efficacy. The recent large core trials have called into question the use of infarct volume alone for selecting patients. This study aimed at exploring the feasibility of using routinely acquired Computed Tomography Perfusion (CTP) to capture how lesion location impacts brain connectivity and how this in turn affects the ability to benefit from EVT. Methods: Consecutive patients from a multi-centric stroke imaging registry were included in this study if they had a proven anterior circulation large vessel occlusion (LVO) and available baseline CTP data. Hypoperfusion (Tmax >6 seconds) and ischemic core (relative cerebral blood flow <30%) maps were automatically generated using e-CTP (Brainomix, UK) and used to derive disconnection maps via a DTI template. A voxel-based, multivariate method (Sparse Canonical-Correlation-Analysis – LESYMAPS) was used to explore associations between disconnection maps and NIHSS subitems in recanalized patients. The mismatch between Tmax and core disconnection volumes (disconnection mismatch volume) (Figure 1) was used to predict modified Rankin Scale (mRS) at 3-months using an ordinal shift regression adjusted for age, recanalization status and infarct volume. Results: One-hundred and seventy-one patients were included. The median NIHSS was 16 (11.5 to 22). The total NIHSS was associated with disconnection predominantly within the left basal ganglia, brainstem and corpus callosum. Upper and lower limb deficits (NIHSS 5a, 5b, 6a, 6b) were associated with corticospinal tract disconnection, with negative correlations within the contralateral tracts (Figure 2). Language deficits were associated with left hemisphere tract disconnection, also showing negative correlations within the contralateral tracts. The disconnection mismatch volume was an independent predictor of functional outcome at 3-months (Table 1). Conclusion: The results from this study showed that it is possible to use routinely acquired CTP to capture the effects of stroke lesion on brain connectivity and its impact on functional outcome following recanalization treatment. Such an approach could enable more nuanced treatment selection for patients based on specific neurological deficits.

Background: Acute stroke lesions become more hypodense on non-contrast CT as the ischemic injury matures. Lesions transition to established infarcts at which time recanalization treatment is thought to be futile, if not harmful. As it is not possible to directly quantify the degree of injury progression, eligibility for treatment is determined using temporal criteria. In this study we used an AI approach to segment and quantify the degree of hypoattenuation of acute stroke lesions on baseline CTs, and explored its impact on clinical outcome following mechanical thrombectomy. Methods: Imaging was collected from a multicentric registry (across 4 US and 5 EU sites) of consecutive large vessel occlusion (LVO) stroke patients who underwent mechanical thrombectomy. Baseline NCCT scans were processed using Brainomix 360 to automatically segment the ischemic core. Lesion hypoattenuation was then quantified by in-house software using the mean Net-Water-Uptake (NWU). Multivariate models were used to explore the association between NWU and clinical outcome (3-month modified Rankin scale–mRS), controlling for e-ASPECTS scores, age, and recanalization status. Interaction models were also used to examine the relationship between NWU e-ASPECTS scores. Results: Of the 1363 patients included (age: median 70, IQR: 50-79; 571 females), successful recanalization (mTICI 2b-3) was achieved for 1119 patients. Multivariate analysis found NWU (p=<.001, t=4.62; Fig1) to remain an independent predictor of outcome. There was a significant interaction effect between and ASPECTS score (p=0.029, Fig 2), suggesting that degree of hypoattenuation played a larger role in patients with lower aspects value. In subgroup analysis, NWU was not a significant predictor of outcome in patients that did not achieve successful recanalizion (mTICI 0-2a). Conclusions: These findings suggest that stroke lesion hypodensity measured using NWU might help identify patients at risk of futile recanalization, and that incorporating this information might be more relevant in patients with lower ASPECTS. The ability to determine treatment eligibility using information extracted from routine imaging would allow safe extension of treatment to patients with unclear onset time, and would account for differences in rates of infarct progression.

Background: There is an unmet need to better understand the likelihood of a good functional outcome in older and frailer populations undergoing endovascular therapy (EVT) following large-vessel occlusion (LVO) ischemic stroke. One approach is to improve the understanding of pre-morbid brain health (“brain frailty”) as a predictor of functional outcome. This study explored the relationship between the spatial pattern of brain atrophy and functional outcome to determine whether this is a predictor of outcomes following EVT. Methods: Patients with LVO from four hospitals in Europe and the USA were included. Three cohorts were used as a derivation cohort, and the fourth served for validation. Presenting CT imaging was processed using Brainomix in-house software to automatically localize and quantify atrophy (frailty-index). Masks from the derivation cohort were used to localize regions of atrophy significantly associated with poorer functional outcome (modified Ranking Scale 3-6) using a voxel-based, multivariate method (Sparse Canonical-Correlation-Analysis) (see Figure 1). For both derivation and validation cohorts, the degrees of overlap between each patient’s atrophy mask and the regions associated at a group level with poor outcome were calculated. This was used as an independent predictor of functional outcome in multivariate analysis adjusting for overall brain frailty-index, recanalization status, and ASPECT score Results: The derivation cohort consisted of 658 patients (mean age 69.6±15.0, median ASPECTS 9 [7-10]), while the validation cohort consisted of 447 patients (mean age 64.7 ±15.5, median ASPECTS 8 [5-10] ). Atrophy localized to the lateral ventricles and sylvian fissure was significantly associated with poorer outcomes (Figure 1). In both derivation and validation cohort, the presence of this spatial pattern of brain frailty (quantified using the degree of overlap) was an independent predictor of poor functional outcome (Table 1). Conclusion: A better understanding of pre-stroke brain health, such as the distribution of atrophy, may provide additional information to inform the prediction of outcome following LVO. This could improve patient selection for treatment following ischemic stroke through better prognostication and a more personalized approach to decision making.

Background: Artificial intelligence (AI) stroke imaging software is becoming mainstay in many hub-and-spoke hospitals. Brainomix 360 Stroke software is the market leader in Europe focused on leveraging simple imaging (non-contrast computerized topography (NCCT) and CT angiography (CTA)) and has recently been FDA-cleared for use in the USA. We evaluated the implementation of Brainomix 360 AI Stroke software in the 17 spoke, multi-state Mayo Clinic Health System (MCHS) telestroke network. Methods: This prospective study compared decision and treatment times before and after Brainomix AI implementation, as well as clinician feedback and simulated decision making to better understand any changes seen. Patients were included if they underwent telestroke evaluation at an MCHS emergency department within 90 days of implementation (2/9/2024-8/8/2024), and excluded if they were already admitted to the hospital or if video evaluation was not performed. Data collection included demographics, clinical decisions and treatment times. Qualitative surveys were conducted at baseline and evaluation end, as well as simulated decision making in 20 de-identified cases randomized to with/without AI through an online portal. Results: A total of 907 patients were included (444 pre- and 463 post-implementation, 287 (32%) with ischemic stroke final diagnosis). Median NIHSS was 2. IVT was recommended in 20.3% (27/148 ischemic stroke patients) pre- and 25.9% (36/139) post-implementation. EVT was recommended in 16.2% (24/148) pre and 14.4% (20/139) post. AI use was associated with trends of faster telestroke activation to IVT decision (36 vs 32 mins, p=0.6), IVT administration (47 vs 40 minutes, p=0.6), and EVT decision (36 vs 33 minutes, p=0.5). In the simulation, imaging interpretation was significantly faster when randomized to AI use (3.4 vs 2.1 mins, p<0.001) and associated with improved physician confidence (62 vs 88%, p=0.004). Qualitative feedback from 14/24 neurologists showed the greatest perceived value was improved confidence of treatment decisions (43%) and faster triaging (36%). Most value was gained from e-ASPECTS (NCCT) and LVO modules. Conclusion: This service evaluation highlighted the opportunity to improve efficiency of decision making even in the setting of a well-developed telestroke system. The impact was underpinned by the improved speed and confidence of imaging interpretation reflected in an offline simulation and qualitative feedback.

Background Identifying eloquent regions associated with poor outcomes based on CT perfusion (CTP) may help inform personalized decisions on selection for endovascular therapy (EVT) in patients with large vessel occlusion (LVO) ischemic stroke. This study aimed to characterize the relationship between CTP-defined hypoperfusion and National Institutes of Health Stroke Scale (NIHSS) subitem deficits. Methods Patients with anterior circulation LVO, baseline CTP, itemized NIHSS at presentation and 24 hours were included. CTP was analyzed using e-CTP (Brainomix, UK). Time to maximal contrast (Tmax) prolongation was defined as >6 s, and penumbra as the difference between Tmax and ischemic core (relative cerebral blood flow<30%). Voxel-lesion-symptom mapping was performed using sparse canonical correlation analysis. For each NIHSS subitem, and total NIHSS, the associations were plotted between Tmax voxels with baseline NIHSS, and penumbra voxels with delta NIHSS (24 hours minus baseline). Results This study included 171 patients. Total NIHSS was predicted by hypoperfusion in left frontal cortex and subcortical white matter tracts. Voxels associated with neurological recovery were symmetrical and subcortical. Limb deficits were associated with respective motor cortex regions and descending motor tracts, with negative correlation within the contralateral hemispheres. A similar but smaller cluster of voxels within the penumbra was associated with NIHSS improvement. Language impairment correlated with left frontal cortex and superior temporal gyrus voxels. With the exception of dysarthria, significant associations were observed and more diffusely distributed in all other NIHSS subitems. Conclusions These results demonstrate the feasibility of hypoperfusion-to-symptom mapping in LVO. Symptom-based mapping from presenting imaging could refine treatment decisions targeting specific neurological deficits.

Background Computed tomography (CT) angiography collateral score (CTA-CS) is an important clinical outcome predictor following mechanical thrombectomy for ischemic stroke with large vessel occlusion (LVO). The present multireader study aimed to evaluate the performance of e-CTA software for automated assistance in CTA-CS scoring. Materials and Methods Brain CTA images of 56 patients with anterior LVO were retrospectively processed. Twelve readers of various clinical training, including junior neuroradiologists, senior neuroradiologists, and neurologists graded collateral flow using visual CTA-CS scale in two sessions separated by a washout period. Reference standard was the consensus of three expert readers. Duration of reading time, inter-rater reliability, and statistical comparison of readers’ performance metrics were analyzed between the e-CTA assisted and unassisted sessions. Results e-CTA assistance resulted in significant increase in mean accuracy (58.6% to 67.5%, p = 0.003), mean F1 score (0.574 to 0.676, p = 0.002), mean precision (58.8% to 68%, p = 0.007), and mean recall (58.7% to 69.9%, p = 0.002), especially with slight filling deficit (CTA-CS 2 and 3). Mean reading time was reduced across all readers (103.4 to 59.7 s, p = 0.001), and inter-rater agreement in CTA-CS assessment was increased (Krippendorff's alpha 0.366 to 0.676). Optimized occlusion laterality detection was also noted with mean accuracy (92.9% to 96.8%, p = 0.009). Conclusion Automated assistance for CTA-CS using e-CTA software provided helpful decision support for readers in terms of improving scoring accuracy and reading efficiency for physicians with a range of experience and training backgrounds and leading to significant improvements in inter-rater agreement.