Automated volume estimate of the brain’s lateral ventricles reveals a persistent association with cognitive performance among stroke survivors

Highlights

• •Longitudinal cognitive trajectories were tracked up to 36 months post-stroke.
• •Automated ventricular volume serves as a macrostructural marker of decline.
• •Log-transformed lateral ventricle volume reliably predicts global MoCA scores.
• •Linear mixed-effects models remain robust when controlling for stroke severity.
• •The study leverages a large, multi-center clinical stroke cohort dataset.

Abstract

Background

Post-stroke cognitive impairment (PSCI) is a significant barrier to recovery. While computed tomography (CT) is the dominant modality for acute stroke diagnosis, it can also quantify brain anatomy, such as lateral ventricular volume (LVV). This study investigates whether initial LVV is associated with Montreal Cognitive Assessment (MoCA) scores over long-term follow-up.

Methods

Initial head CT, longitudinal MoCA, and functional assessments were collected at sites from the Norwegian COgnitive impairment After STroke (Nor-COAST) study. LVV was estimated using an established deep learning segmentation tool. Associations between initial LVV and repeat MoCA were tested across four timepoints (up to three years) using a linear mixed effects model. To provide context, we explored the independent influence of LVV in relation to National Institutes of Health Stroke Scale (NIHSS), modified Rankin Scale (mRS), and Global Deterioration Scale (GDS).

Results

The sample of N = 547 participants were 74 ± 12 years (range: 35–97 years) and 54% female. Log-transformed LVV was inversely associated with longitudinal MoCA scores (t=-3.75, p < 0.001). This relationship was consistent across cross-sectional analyses at individual time points. Secondary analyses revealed significant LVV associations with NIHSS, mRS, and GDS scores (all p < 0.05).

Conclusion

LVV was independently associated with lower MoCA scores across the three-year follow-up after stroke. Specifically, a patient with twice the initial LVV scored 0.7 points lower on the MoCA across three years of follow-up. These findings reveal that CT obtained at stroke admission may yield prognostic information about long-term cognitive outcomes, supporting the case for automated LVV estimation.

Introduction

Stroke is a leading cause of disability and death, with an estimated 12 million strokes occurring annually worldwide [1]. Acute stroke diagnosis and treatment rely on clinical assessment (i.e. motor, facial, sensory, and language deficits) and imaging to determine aetiology and lesion location [[2], [3], [4]]. There is, however, increased recognition of the high prevalence of cognitive impairment at all stages after stroke [5]. Post-stroke cognitive impairment (PSCI) is defined as a cognitive deficit that develops in the first three months following stroke, persists for a minimum of six months, cannot be explained by another condition, and involves language, executive function, visuospatial, episodic, or working memory cognitive domains [6,7]. PSCI is highly relevant to stroke recovery and portends a significant dementia risk, as reflected by 59% of patients with cognitive impairment at three months[8], and approximately 33% of patients with PSCI that convert to dementia in the years post-stroke [2,9,10].

Research on screening to PSCI is warranted, particularly exploiting imaging where biomarkers of ageing and cognition are well established. Indeed, comprehensive cognitive work-up in the early stages after stroke and during recovery is resource intensive. It requires dedicated personnel and may burden patients and their caregivers [11,12]. Non-contrast Computed Tomography (CT) imaging is ubiquitous in the acute stroke workup. CT images contain morphological and anatomical details that may relate to cognition, in addition to the stroke related imaging information [[13], [14], [15]]. For instance, the brain’s ventricles are easily identifiable on CT [16], providing the means for segmentation. Although ventricle volume is not reported as part of standard acute stroke care, ventricular enlargement and/or brain tissue atrophy [17,16] are independent markers of brain health, cognitive decline, and long-term outcomes after stroke [14,18,19]. Quantification of the lateral ventricle volume (LVV) is not currently used to inform on PSCI risk despite an association that relates to neurodegeneration processes [13]. Hence there is potential clinical value in fully automated segmentation methods to estimate ventricle volumes on CT [19].

The current study focuses on LVV in relation to cognitive, functional, and stroke severity outcomes. The Norwegian COgnitive impairment After STroke (Nor-COAST) study offers data that are uniquely suited for the current investigation, as it involved baseline imaging and up to three years of longitudinal assessments among stroke survivors [20]. The primary objective was to test for associations between LVV assessed at acute presentation and the Montreal Cognitive Assessment (MoCA) score over time. The study also aims to investigate cross-sectional relationships between the LVV and other cognitive, clinical, and stroke impairment scales. To facilitate the LVV estimate we deployed an automated deep learning tool that segments the lateral ventricles, along with any type of intracranial hemorrhage (called ventricle and bleed artificial intelligence (AI), VB-AI), which was designed for use in both stroke and traumatic brain injury [15].

Methods

Participants

Inclusion was based on participation in the Nor-COAST study which recruited stroke patients across five sites in Norway. It was a prospective study aimed to recruit patients at distinct phases after stroke, including hospital admission as baseline (T0), 3 months (T1), 18 months (T2), and 36 months (T3). Data collection occurred between 2015 and 2020 and was performed according to the Helsinki declaration [21]. Participants provided informed written consent. Among individuals who were unable to provide consent, proxy consent was provided by a family member. Nor-COAST was approved by the regional ethics committees for medical and health research (REK Nord, number: 2015/171 & 2016/2306).

The following criteria were used to create the current study sample: Individuals were included if they were admitted to hospital with acute ischemic or haemorrhagic stroke, with age above 18 years, fluent in a Scandinavian language, availability of a routine non-contrast head CT, and completion of at least one cognitive assessment (described below). Exclusion criteria were diagnosis other than stroke, and lack of CT, demographic or medical details. Cohort construction also relied on accessing information that included medical history, attained years of education, and smoking and marital status. Inspection of the baseline CT and cognitive data was performed to create the cohort.

Assessments

Clinical, functional, and cognitive assessments were performed in-hospital, at the outpatient clinic, or over the phone at given time intervals post-stroke, within the time frames provided above (i.e. T0 to T3). Assessments were selected while considering the recommendations from the National Institute of Neurological Disorders and Stroke Canadian Stroke Network (NINDS-CSN) [22]. The list of assessments included the MoCA, screening for cognitive function using the GDS, stroke severity by the NIHSS, and pre- and post-stroke function measured by the mRS.

The MoCA is a 30-point cognitive screening test validated for use after stroke [23,24], with domains that include memory, visuospatial abilities, executive functions, attention, language, and orientation to time and place. MoCA was performed by a clinician and assessed for up to 36 months post stroke at four assessment times (i.e. T0 to T3). The 11-item NIHSS neurological exam (score range: 0 to 40) was performed at initial presentation. The NIHSS served as a proxy for overall stroke deficits (motor and sensory impairments, language, and visual), with a low score denoting lower stroke severity [25], and is relevant when considering cognitive status after stroke [26]. The mRS is a widely used scale to assess independence of daily activities (mRS of 0 is no disability, 5 is disability requiring constant care for all needs, and 6 is death) [27,28]. The mRS was assessed as pre-stroke based on recollection by the participant or with support from a family member, at three-months (T1), and at eighteen-months (T3). GDS was collected by a clinician at baseline (T0) in an interview with the participant and/or next of kin, using a seven-point scale to evaluate subjective and objective memory deficits [29,30].

Imaging

Routine CT images were acquired at participating sites at the time of the initial acute stroke presentation (T0). Images were reconstructed with an average voxel resolution of 0.42 × 0.42 × 2.84 mm³ and 55 ± 15 axial slices (see Fig. 1), covering the whole brain. LVV and total intracranial volumes were extracted from the CT scans using automated methods [15,31]. LVV quantification used an in-house approach referred to as “VB-AI” and uses a dual segmentation deep learning method to segment the left and right lateral ventricles along with a second segmentation of intracranial hemorrhage.

Fig. 1.

Two cases are shown in A) and B) with CT images (greyscale). The VB-AI based LVV segmentation (blue) is overlaid. C) The distribution of LVV values for the cohort is provided and shown as a histogram prior to log-transform.

The architecture for VB-AI was an ensemble of no-new U-net (nnU-net) architecture models [32], trained with cross-validation through five data folds. As described in more detail elsewhere, a self-configuring 3-dimensional nnU-Net model (version 2 and using the 3d_fullres option) was chosen after considering four other deep learning configurations [15]. In brief, the training data consisted of n = 154 cases of raw CT images and corresponding annotations (69 cases from the open-source INSTANCE 2022 challenge, 65 cases and 20 cases from two local hospitals in the Oslo, Norway). The models were trained for 1000 epochs using an RTX 3090 24GB graphics processing unit (GPU), batch size of 2, patch size of [160,160,96], over approximately 110 h. No manual preprocessing was applied. Images were handled using nnU-Net’s default self-configured preprocessing. Hence, the lateral ventricle segmentation method was developed on images entirely distinct from the current study, noting that the external CT were acquired with comparable acquisition details, diverse cases, and a range of voxel dimensions and the number of slices [33]. VB-AI was able to expedite LVV quantification, which would otherwise be labour intensive. Visual inspection (BJM, PM, EM) revealed segmentation fidelity that demarcated the boundaries of the normal lateral ventricles anatomy, making this approach suitable to estimate LVV. We used the same GPU for training (i.e. external data) and inference images (i.e. Nor-COAST data). A docker executable version of the VB-AI software tool is available for research use through requests sent to the corresponding author.

Our own work and that of others reveals the relevance of accounting for age, sex, and head size (i.e. intracranial volume) when considered LVV estimates, which motivated the need to include age and sex in all statistical models that test for an association between a cognitive or functional score and LVV [34].

Statistics

We performed two sets of analyses to investigate the primary (i.e. MoCA) and secondary (NIHSS, mRS, GDS) study aims. Through data inspection, there was recognition that the raw LVV estimates showed a right skewed distribution, which was formally evaluated using Shapiro-Wilk test. As a result, the LVV estimate was log-transformed using the natural logarithm to yield a normal distribution for this independent variable in all subsequence regression models. Descriptive statistics were compiled for the sample to yield demographics information.

To test the primary hypothesis, we used a linear mixed effects model. The independent explanatory variables in this model included log-LVV, timepoints (i.e. up to a max of four), sex, age, years of education, initial NIHSS, intracranial volume, and a log-LVV by timepoint interaction term. Participant study ID and imaging site were included as random effects. The interaction term was incorporated to account for potential sources of bias in the participants with fewer versus more repeat timepoints. Independent variables were assessed for multi-collinearity. Initial evaluation revealed the variance inflation factor was < 3, which is deemed low risk for variance inflation for the regression model parameters.

To provide context for the linear mixed effects results, separate cross-sectional linear regression models were used to examine the association between MoCA performance at a given timepoint post-stroke and log-LVV, while accounting for covariates. This same cross-sectional regression approach was used for each of the functional outcome measures that were considered for the secondary analyses. Namely, we also investigated whether there was an independent influence of log-LVV on the following outcome measures: initial NIHSS, pre-stroke mRS, 3-months mRS, 18-months mRS, and GDS. As reported below, the overall sample size tended to differ for these analyses due to a small number of missing data, however, the independent variables remained identical for each regression test (i.e. log-LVV, age, sex, education, intracranial volume, and imaging site).

For all models, the raw and standardized beta coefficients were calculated for the regression models, along with standard errors and unadjusted p-values, and the adjusted model R-squared values. An unadjusted p-value < 0.05 was considered significant. The R programming language [35] was used for the calculations using the lm4 [36], nlme [37], car and the dplyr [38] packages.

Results

A total of 547 stroke survivors from the Nor-COAST study population met the current study inclusion criteria. Table 1 shows the demographic and clinical details, which shows that the sample was nearly balanced for sex with a slightly higher percentage of females. The sample consisted of 89% with ischemic stroke, of which 22% of patients that had received thrombolysis and 2.8% that received thrombectomy (Table 1). Table 2 shows the summary values for the NIHSS, mRS, and GDS assessments.

Table 1.

Demographic and clinical details are provided for the study participants. Data are reported based on N = 547 participants that was the largest sample size available for cross-sectional analysis. Diabetes was based on HbA1c ≥48 mmol/mol or use of antidiabetic drugs at baseline. (The * indicates there were between 1 and 9 missing cases for the specific variable.).

Characteristics	Mean (SD) or N(%)
Age (Years)	74 (12)
Sex*
Male	250 (46%)
Female	288 (54%)
Education (Years)	12.2 (3.7)
Intracranial Volume (mL)	1430 (159)
Marital Status*
Married	326 (61%)
Single	82 (15%)
Divorced/Widowed	130 (24%)
Smoking History*
No	231 (43%)
Yes	87 (16%)
Previously	217 (40%)
Diabetes	103 (19%)
Thrombolysis*	118 (22%)
Thrombectomy	15 (2.8%)
Stroke Type
Ischemic	474 (89%)
Hemorrhagic	61 (11%)

Table 2.

Summary of outcome variables used in the primary or additional analyses. Abbreviations: Global Deterioration Scale (GDS), modified Rankin Scale (mRS), Montreal Cognitive Assessment (MoCA).

Assessment	Summary
MoCA (Baseline)	Mean: 23 (SD: 6) | Median: 25 [IQR: 20–27]
NIHSS (Initial)	Mean: 5 (SD: 6) | Median: 3 [IQR: 1–6]
mRS (3-months)
0	48 (10%)
1	149 (32%)
2	149 (32%)
3	57 (12%)
4	45 (9.6%)
5	20 (4.3%)
GDS (Baseline)
1	362 (67%)
2	83 (15%)
3	45 (8.4%)
4	26 (4.8%)
5	17 (3.2%)
6	5 (0.9%)

The linear mixed effects model relied on 1452 MoCA scores from among n = 482 unique participants. We found that MoCA performance over time was significantly associated with log-LVV (B=−1.35, SE=0.36, t=−3.75, p < 0.001), age (B=−0.15, SE=0.02, t=−7.43, p < 0.001), education (B = 0.42, SE=0.06, t = 7.47, p < 0.001), and initial stroke severity (NIHSS: B=−0.20, SE=0.04, t=−4.90, p < 0.001). There was no significant main effect of timepoint (B = 0.42, SE=0.38, t = 1.11, p = 0.269) and no significant log-LVV by timepoint interaction (B=−0.16, SE=0.11, t=−1.43, p = 0.153). MoCA performance was not significantly related to sex (B=−0.63, SE=0.49, t=−1.29, p = 0.197) or total intracranial volume (B = 0.002, SE=0.002, t = 1.10, p = 0.271). In support of the primary aim, we examined regression models for the individual MoCA assessments and found that log-LVV remained significant at the individual timepoint level (p < 0.016) (Fig. 2A; results of the individual MoCA regression models are provided in a Supplementary Table).

Fig. 2.

Regression results are shown between explanatory variables (i.e. log-LVV) and outcomes, reported as standardized beta coefficients (with 95% confidence intervals) and based on baseline (T0), 3 months (T1), 18 months (T2), and 36 months (T3) timepoints. A) Longitudinal associations between MoCA score and LVV at each of the four assessment timepoints. Each of the MoCA timepoints showed a standardized beta that was negative, reflecting a significant inverse influence for LVV. MoCA associations at T1 are also shown for age, years of education (Edu), and NIHSS model parameters. B) Additional results for LVV are shown for secondary outcomes. Aside from NIHSS vs LVV which showed a modest negative association, all other of the relationships were positive associations with LVV. Abbreviations: Lateral ventricle volume (LVV), Montreal Cognitive Assessment (MoCA).

The results of the additional cross-sectional linear regression models are reported as the following and for which all models accounting for intracranial volume and imaging site (Fig. 2B). The initial GDS (n = 538) was significantly associated with log-LVV (B = 0.33, SE=0.075, t = 4.38, p < 0.001), age (B = 0.022, SE=0.0045, t = 4.93, p < 0.001), and education (B=−0.036, SE=0.013, t=−2.73, p = 0.006), but not sex (B = 0.006, SE=0.109, t = 0.06, p = 0.95). The initial NIHSS score (N = 547) was negatively associated with log-LVV (B=−1.00, SE=0.46, t=−2.17, p = 0.031) and education (B=−0.25, SE=0.08, t=−3.11, p = 0.002), with a non-significant trend for age (B = 0.05, SE=0.03, t = 1.92, p = 0.056), but not significantly associated to intracranial volume (B = 0.002, SE=0.002, t = 0.86, p = 0.393). The prestroke mRS score (N = 547) was associated with log-LVV (B = 0.33, SE=0.071, t = 4.58, p < 0.001), age (B = 0.029, SE=0.004, t = 6.65, p < 0.001), and sex (B=−0.22, SE=0.105, t=−2.09, p = 0.037), but not to years of education (B=−0.016, SE=0.013, t=−1.25, p = 0.214). The 3-months mRS score (n = 468) was associated with log-LVV (B = 0.19, SE=0.090, t = 2.15, p = 0.032), age (B = 0.027, SE=0.005, t = 4.96, p < 0.001), and sex (B=−0.39, SE=0.132, t=−2.96, p = 0.003), but not education (B=−0.023, SE=0.015, t=−1.48, p = 0.140). The 18-months mRS score (n = 397) was associated with log-LVV (B = 0.49, SE=0.099, t = 4.93, p < 0.001), age (B = 0.023, SE=0.006, t = 3.80, p < 0.001), and education (B=−0.040, SE=0.017, t=−2.33, p = 0.021), but not sex (B=−0.21, SE=0.146, t=−1.45, p = 0.147).

Discussion

The current study reveals a persistent relationship between acute stroke imaging lateral ventricular volume (LVV) and MoCA performance up to 36 months post-stroke. By estimating LVV from non-contrast CT, we found that larger baseline LVV was independently associated with lower MoCA scores, a relationship that was stable over the multiple cognitive screening assessment timepoints. The primary aim was supported by a series of additional analysis, which helped to demonstrate that larger initial LVV was also related to measures of functional and overall impairment. Collectively, these data demonstrate that quantification of brain anatomy from the acute stroke CT can provide relevant and significant context for the stroke patient throughout the acute to chronic stages of recovery.

The persistence of the LVV-MoCA association is likely a reflection of underlying factors that contribute to inter-individual differences, such as resilience to developing PSCI after stroke. Although MoCA is a screening tool designed to capture multi-domain cognitive impairment, the findings of this study suggest there is sensitivity to underlying structural brain health, specifically ventricle size. Importantly, this relationship remained stable even as patients transitioned from the acute to the chronic recovery phases. Other studies, including our own work, have reported on the temporal change in MoCA performance as a function of time post-stroke [39]. For instance, a recent study demonstrated an associations between follow-up MoCA performance and the acute structural T1-weighted MRI features [40]. Previously, we also found MRI-based metrics that were associated with cognition after stroke, which when combined with clinical variables was able to fairly accurately classify individuals based on future risk of cognitive disorder within months after stroke [41]. Another longitudinal stroke study revealed greater reductions in normal-appearing white matter integrity over time was associated with less improvement on MoCA performance [42]. Collectively, the current study adds to the prior research in demonstrating the relevance of brain imaging markers in PSCI.

Additional analysis involving the mRS, GDS, and NIHSS provided important context for the influence of LVV across a spectrum of post-stroke outcomes. Regarding the mRS, we found that higher LVV was consistently associated with greater functional disability—both based on the pre-stroke patient report and at 3- and 18-months post-stroke. This positive association aligns with the established role of ventricular enlargement as a proxy for neurodegeneration. Notably, the association between LVV and mRS reached its highest magnitude at the 18-month mark. This suggests that while acute functional status may be dictated by the stroke lesion itself, imaging measures that reflect underlying structural integrity are relevant in the chronic phase, as was reflected in the current study by large standardized beta coefficients (a proxy for effect size) for each of the timepoint considered. Relatedly, LVV was also positively associated with the Global Deterioration Scale (GDS), which again reflects larger ventricles as a marker of neurodegeneration that may have preceded the stroke and/or persist thereafter even for mild to moderate stroke. It is important to highlight that the negative association between LVV and initial NIHSS was somewhat counterintuitive. This association should be viewed with caution for a few reasons. The explained variance for this regression model was weak and might have been influenced by the relatively low and narrow range of NIHSS values in this sample. Also, we did not account for stroke type (i.e. ischaemic, lacunar, or hemorrhagic stroke) or the time between stroke onset and NIHSS assessment, which could include the NIHSS finding and therefore future research is warranted.

Several limitations of this study warrant consideration. First, potential selection bias and retention bias may limit the generalizability of our findings, as our sample predominantly comprised patients with relatively low baseline NIHSS scores. Furthermore, we did not include participants who were unable to complete the MoCA, which was part of the rationale for incorporating the GDS in the current study. Second, the MoCA is not immune to practice effects; serial testing over time can induce a training effect that tends to artificially elevate subsequent scores [43]. Third, we did not consider the minimum clinically important difference (MCID) for the MoCA in our primary analyses [44]. We attempted to account for this in the MoCA linear mixed effects model by including timepoint as a fixed effect and furthermore, the interaction of timepoint and LVV was not significant, which provide assurance for the current findings. Relatedly, retention bias is a factor in a longitudinal study. The MoCA analysis did account for NIHSS, which was significant as expected, and can be viewed as a strength of the chosen statistical design. We did not, however, incorporate death or dementia as competing outcomes and this would be interesting to consider in future work. Lastly, we did not include imaging-derived metrics, such as stroke volume or specific lesion location, within the regression models because our explicit intent was to focus on the LVV imaging marker. Analyses did account for total intracranial volume (i.e. head size) as this was relatively simple to calculate from the CT images. This imaging measure does not reflect the brain parenchymal fraction, which is more challenging to calculate with CT but would be interesting to consider as another source of between-participant differences. Repeat CT imaging was not part of the study, and this would be interesting to monitor LVV from among other imaging-derived estimates that may change over time.

From a clinical workflow perspective, the results of the current study reinforce the potential for automated and quantitative LVV estimates using standard of care CT that is typically the first image collected as part of a stroke imaging workup. The VB-AI tool is an example of the growing use of deep learning software to augment stroke imaging research, among other CT image analysis software solutions [45,46]. Indeed, there are comprehensive stroke imaging AI studies (e.g. classification, segmentation, or outcome prediction) [47], however, the current study is a novel contribution to the literature as there are no studies that focus on CT-based LVV in relation to PSCI and stroke outcomes [48]. From the current study we extracted an objective and structural brain health measure to gain insight into a patient’s long-term cognitive and functional prognosis without the need for additional imaging. Although MRI is superior to CT for soft tissue contrast, head CT is clinically dominant [49]. Our findings reveal that a CT-based metric offered critical information regarding the persistently lower cognitive performance on a screening test. This research highlights the importance of identifying markers that increase the risk of PSCI, with treatment that may be better timed or personalized to through characterization of the inherent heterogeneity among stroke survivors.

Conclusion

This study demonstrates that an automated estimate of LVV, combined with stroke related parameters, served as a robust and persistent imaging marker for post-stroke cognitive performance. The results indicate that it is both viable and useful to quantify LVV from standard-of-care acute head CT scan and may be useful to characterize cognitive trajectories up to three years post-stroke. Ultimately, leveraging AI-driven metrics like the LVV in the acute setting represents important steps toward rapid brain health phenotyping, the potential for more personalized stroke recovery strategies, and more efficient resource allocation within clinical pathways.

CRediT authorship contribution statement

Bradley J MacIntosh: Writing – original draft, Visualization, Software, Methodology, Formal analysis, Data curation, Conceptualization. Pål CW Morberg: Writing – review & editing, Writing – original draft, Formal analysis, Data curation, Conceptualization. Elakkyen Murugesu: Writing – review & editing, Software, Formal analysis, Data curation. Mona K Beyer: Writing – review & editing, Project administration, Investigation, Funding acquisition, Data curation, Conceptualization. Karianne Larsen: Writing – review & editing, Methodology, Formal analysis. Karoline Skogen: Writing – review & editing, Methodology, Formal analysis, Data curation. Eva B Aamodt: Writing – review & editing, Methodology, Formal analysis, Data curation, Conceptualization. Atle Bjørnerud: Writing – review & editing, Software, Methodology, Data curation. Inge R Groote: Writing – review & editing, Software, Methodology, Data curation. Matthan WA Caan: Writing – review & editing, Software, Data curation. Hege Ihle-Hansen: Writing – review & editing, Methodology, Investigation, Formal analysis, Data curation. Ingvild Saltvedt: Writing – review & editing, Project administration, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Till Schellhorn: Writing – review & editing, Software, Methodology, Formal analysis, Data curation, Conceptualization.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Bradley MacIntosh reports a relationship with Antaros Medical Inc that includes: consulting or advisory. Bradley MacIntosh has patent pending to Licensee. Other authors have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.