Automated Quantitative Assessment of Cerebral Edema after Ischemic Stroke using CSF Volumetrics

Abstract

Reduction in CSF volume from baseline to follow-up CT at or beyond 24-hours can serve as a quantitative biomarker of cerebral edema after stroke. We have demonstrated that assessment of CSF displacement reflects edema metrics such as lesion volume, midline shift, and neurologic deterioration. We have also developed a neural network-based image segmentation algorithm that can automatically measure CSF volume on serial CT scans from stroke patients. We have integrated this algorithm into an image processing pipeline that can extract this edema biomarker from large cohorts of stroke patients. Finally, we have created a stroke repository that can archive and process images from thousands of stroke patients in order to measure CSF volumetrics. We plan on applying this metric as a quantitative endophenotype of cerebral edema to facilitate early prediction of clinical deterioration as well as large-scale genetic studies.

The Spectrum of Cerebral Edema after Stroke

Stroke is a leading cause of death globally, responsible for an estimated six million deaths each year, as well as an enormous burden of disability [1]. While ischemia and ensuing cerebral infarction result in neurologic deficits, it is not the infarct itself that is responsible for most of stroke deaths in the acute period. Instead, a majority of death and neurologic deterioration after hemispheric infarction is related to the development of brain swelling (i.e. the pathologic accumulation of excess water around infarcted tissue) [2]. This process leads to mass effect and deterioration (i.e. malignant edema) in 10–30% of patients with hemispheric strokes and accounts for half of all hospital deaths [3]. Furthermore, ischemic edema results in secondary brain injury and may jeopardize neurologic recovery even in those without malignant edema [4]. Greater volume of edema (assessed using serial MR imaging) predicts poor outcome, independent of infarct volume and stroke severity (e.g. NIHSS) [5].

Although malignant cerebral edema is a well-recognized, life-threatening complication of stroke, scientific study of brain swelling remains rudimentary. Current clinical assessments to detect the development of edema either require close bedside monitoring for neurologic deterioration or radiologic evaluation of crude swelling biomarkers such as lesion volume (a variable combination of infarct plus edematous tissue) or midline shift. Although midline shift is associated with reduced level of consciousness [6], it only develops relatively late in the edema cascade, when room to compensate for increased brain volume has been exhausted (primarily by displacement of blood and cerebrospinal fluid [CSF]) [7]. Measuring midline shift as a biomarker of edema also only provides a partial picture of the spectrum of swelling that occurs, as many stroke patients with moderate degrees of edema will never develop midline shift; the process of edema could still be relevant to their neurologic outcome even in the absence of visible shift and clear deterioration in mental status [5]. MRI-based measures that quantify edema development around an infarct lesion have been developed but are labor-intensive and are best suited to larger lesions imaged once swelling is already established [5].

Furthermore, waiting for swelling and midline shift to develop exposes the stroke patient to the risk of cerebral herniation and secondary injury that may not be reversible. Being able to recognize brain swelling earlier and predict those at risk of deterioration would enable initiation of targeted therapies to minimize edema and selection of candidates for aggressive surgical procedures (i.e. decompressive hemicraniectomy) [8]. Currently, early radiographic assessment of edema requires MRI-based evaluation of lesion volume [9]. This may not be feasible or cost-effective for all patients at all stroke centers worldwide. Instead, computed tomography (CT) imaging is routinely performed at presentation and often for follow-up (FU) of stroke patients [10]. However, CT is insensitive to measuring lesion or edema volume, especially early after stroke.

The need for better tools to measure edema early using standard CT has led to the exploration of two differing approaches (see Table 1 for a comparison of various methods to quantify edema after stroke). One estimates the volumetric brain water increase within ischemic tissue using an imaging assessment of “net water uptake” (NWU) [11]. This employs a densitometric analysis of the ischemic “lesion” compared to a mirrored region in the contralateral normal hemisphere. Measurements of NWU have shown promise in distinguishing early infarct lesions destined for malignant edema from others with large-vessel occlusion without significant edema [12]. However, this approach requires either the region of infarction to be visible (not commonly seen on CT performed within 12–24 hours of stroke) or an estimate of infarct core from concomitant CT perfusion. Additional studies have demonstrated that change in NWU over time (as a surrogate of edema) is greater in those without successful recanalization [13]. However, NWU cannot be calculated in those with hemorrhagic transformation or without visible infarct, thereby limiting its broad applicability in both large strokes with hemorrhage and smaller strokes.

Table 1:

Comparison of methods for quantifying brain edema after ischemic stroke

	Infarct Volume	Midline Shift	MRI-based FLAIR	CSF Volumetrics	Lesional Water Uptake
Captures	Infarct and associated swelling	When edema is severe enough to exhaust reserve	Edema around stroke [5]	Displacement of CSF in compensation for increased brain volume	Increased brain water within lesion
Imaging	MRI > CT	CT or MRI	MRI × 2	CT × 2	CT
Timing	Early with DWI MRI Later with CT	Develops late	At baseline and peak edema	Within 24 hours (comparison with baseline CT)	Early (within 6–12 hours) requires CTP
Ability to quantify full spectrum of edema and kinetics	Yes, but not separate infarct from swelling	Only late stages	Yes	Yes	Yes (only if infarct visible)
Associated with deterioration	Yes	+++	+	+	+
Accurate in selecting for DHC early	Yes (80–90%)	Too Late	Unknown	Yes (preliminary)	Yes (preliminary)
Ease of measuring	Manual tracing of lesion	Relatively simple	Manual tracing of lesions	Time consuming to outline CSF	Manual tracing of lesion and midline
Automated quantification	+/−	+/−	No	Yes	+/−

CT, computed tomography; CTP, CT perfusion; DHC, decompressive hemicraniectomy; DWI, diffusion-weighted imaging; MRI, magnetic resonance imaging

We have proposed that measuring displacement of CSF from the cerebral hemispheres (both sulci and ventricles) between baseline and FU CT scans can provide a quantitative imaging biomarker of cerebral edema after ischemic stroke [14]. We will outline the rationale and results supporting this approach as well as suggest that CSF volumetrics can provide insights into early edema formation (i.e. within the first 24-hours after stroke), before significant midline shift has developed. Furthermore, we will describe how this biomarker can be leveraged (using an automated imaging pipeline) for large-scale studies of edema that could uncover new insights into the time course of edema and molecular targets for stroke and edema therapeutics.

Rationale for Using CSF Displacement as a Biomarker of Cerebral Edema

Total cranial volume inside the closed skull is fixed and comprises the sum of brain, blood, and CSF [15]. This Monro-Kellie doctrine states that any increase in one of these compartments (for example, brain swelling) must result in a reciprocal compensatory decrease in other compartments (i.e. reduction in intracerebral blood volume and displacement of CSF from the cranial vault). In fact, Harvey Cushing viewed the CSF as the brain’s “third circulation” with a critical role in such compensatory responses that mitigate increases in intracranial pressure from mass lesions [16]. The baseline volume of cranial CSF may provide a reserve for swelling, with those (often younger) patients with lower CSF volumes being at higher risk of malignant edema [17]. It follows that there will be a decrease in hemispheric CSF volume as edema forms around the evolving infarct and brain volume increases. This CSF displacement (from sulci and eventually ventricles) should be reciprocally related to the severity of swelling; a relationship that can be leveraged to indirectly measure the volume of edema (i.e. brain volume increase) that is otherwise inaccessible. Furthermore, the CSF compartment can be visualized on both MRI and more routinely available CT scans, making it a more accessible imaging biomarker.

We proposed that the change in CSF volume (ΔCSF) between baseline and FU CT scans can serve as a quantitative biomarker of cerebral edema after stroke. In a pilot cohort of patients with large hemispheric infarction, we demonstrated that ΔCSF (from baseline to FU at time of peak edema, usually 48–72 hours after stroke) was correlated with extent of midline shift (r=0.75) [14]. While the majority of ΔCSF (mean of 55-ml in total) was lost from ipsilateral sulci (29-ml), there was also CSF displaced from the contralateral sulci (14-ml) and ipsilateral ventricle (10-ml). Furthermore, extent of ΔCSF was able to distinguish those with malignant edema, defined as those requiring osmotic therapy or decompressive hemicraniectomy from those with similar stroke severities who did not develop malignant edema.

In addition, we observed that over half this CSF displacement was already seen on scans performed 24 hours after stroke, suggesting that ΔCSF better reflects the acute effects of edema than later measures such as midline shift. In Figure 1, serial CTs of a young stroke patient are displayed, demonstrating that segmented CSF volume is reduced by 43% (from 102 to 58-ml) by the time of early follow-up CT performed within 24 hours of stroke, before midline shift or even clear stroke-related hypodensity are visible. By the time midline shift and large hypodensity are seen at 36-hours after onset, there is further reduction in ventricular and sulcal CSF. It was only at this point that malignant edema was diagnosed and the patient subsequently underwent hemicraniectomy. However, recognition of rapid early CSF volume reduction might permit triage of such patients to surgery prior to deterioration.

Figure 1:

Serial CT scans from a 28-year old woman with uncontrolled diabetes mellitus who presented within one hour of the onset of symptoms compatible with right middle cerebral artery (MCA) stroke with baseline NIHSS of 13 and glucose over 400 mg/dl. She received tPA after the baseline CT (A) was unrevealing. Baseline CSF volume was measured to be 112-ml. She underwent repeat CT (B) the next day (14 hours after stroke onset) with early changes of MCA infarction but no significant midline shift. CSF volume was already reduced to 58-ml (i.e. by almost half). Her neurological status worsened (NIHSS 17) and she underwent a repeat CT (C) at 35-hours after stroke. This revealed large right MCA hypodensity (over 300-ml) with 6-mm of midline shift. CSF volume was further reduced to 38-ml.

The concept that CSF volume reduction can be used as a surrogate of edema volume has been further validated in a recent study where CSF was segmented from baseline and 5-day MRI [18]. ΔCSF was shown to parallel a reference measure of swelling, i.e. the difference between 5-day and 90-day lesion volume on FLAIR imaging. There have also been other studies where CSF volumetrics have been applied as meaningful dynamic measures of edema, for example, in liver failure [19]. The reduction in sulcal CSF volume has also been proposed as a quantitative biomarker of edema in subarachnoid hemorrhage [20].

Automated Measurement of CSF Volumes Using Machine Learning

In our pilot evaluation confirming the relationship of ΔCSF and edema, the volume of CSF was obtained by time-consuming manual outlining of all sulci and ventricular regions on serial CT scans. This approach is clearly infeasible if this biomarker is to be applied to understand edema across larger cohorts of stroke patients. For that reason, we have developed a computer algorithm capable of automatically segmenting CSF from standard clinical CT images [21]. To accomplish this we employed a supervised machine-learning approach where a random forest model was trained on manually delineated CSF regions from paired (baseline and FU) CT scans from 38 stroke patients from two centers. The resulting algorithm was able to measure CSF volumes with high concordance to manually obtained measurements (r=0.95). We have further refined this algorithm by training a fully convolutional neural network (based on the U-Net architecture [22]) to even more accurately perform the segmentation of CSF. The segmentations shown in Figure 1 were performed by the automated deep learning algorithm on clinical CTs and take less than one minute per scan.

Image Processing Pipeline

We have also developed a series of image processing steps to ensure that consistent cranial volume is extracted from CT scans (that each may have variable brain coverage) and that serial CT scans from the same patient are co-registered to optimize stability of volumetric measurements. Integration of these pre-processing steps with the segmentation algorithm allows us to obtain ΔCSF between scans at several time points after stroke rapidly and without manual intervention [23]. This pre-processing of serial scans for a single patient takes between one and two hours on a single processor but can be parallelized on a cluster (see below). The only tasks that still require human input are the measurement of midline shift (as a comparator end-point) and delineation of the infarct-related hypodensity region on CT. We are currently in the process of automating both of these steps to allow end-to-end analysis of multiple stroke-related imaging phenotypes in a single pipeline.

In order to facilitate neuroimaging analyses on a large scale, we have created a stroke-focused imaging repository based on the XNAT (The Extensible Neuroimaging Archive Toolkit) informatics platform [24]. We have labeled this: the Stroke NeuroImaging Phenotype Repository (SNIPR). SNIPR already contains brain imaging on over 2,000 stroke patients from several international sites. XNAT also has the ability to run Docker containers on the imaging data. This open-source tool provides a means of loading modules of the processing pipeline into standardized self-contained containers and running them on the stroke data to extract relevant phenotypes such as CSF volume. We have now transformed all the steps of our edema pipeline into a series of containerized modules that can run on all the stroke images in SNIPR. We have also added modules that perform quality control and ensure the required scan type (i.e. axial non-contrast brain sequence for CSF analysis) is available at each time point (an outline of the pipeline shown in Figure 2). We are now developing further quality checks that can automatically detect poor quality CTs (e.g. with significant motion artifact) that may be unable to be utilized for such analyses. This platform also allows parallel processing of many patients’ scans on a high-performance computing cluster to greatly increase the throughput of the pipeline.

Figure 2:

Image processing pipeline built into SNIPR to extract CSF volumes from serial CT scans of stroke patients. Anonymized DICOM (Digital Imaging and Communications in Medicine, DCM) image files are stored in the repository. The pipeline selects which imaging sequences at each time point represent axial brain scans (i.e. excluding bone scans, sagittal and coronal reconstructions) and converts these to NIfTI (Neuroimaging Informatics Technology Initiative, NII) research files. These files are further processed to extract brain regions (i.e. skull removal) and register each scan from a given patient to one another. These registered images are then segmented into brain and CSF and CSF volume calculated after exclusion of regions with visible infarct.

Applications of ΔCSF to Better Understand Stroke Edema

We have demonstrated that ΔCSF (change in CSF volume from stroke presentation to 24-hour or beyond) can serve as a biomarker of brain swelling. It is correlated with subsequent infarct volume and midline shift, but at time points before either of these two are easily measurable on CT. Furthermore, it provides a quantitative measure of edema that can be derived from routinely available CT scans. We have developed a means of measuring ΔCSF using a deep learning-based segmentation algorithm. We have a scalable means of extracting this biomarker in thousands of patients using a centralized imaging repository and containerized processing modules. This now allows us to better study the kinetics of edema across large stroke cohorts and facilitate the following avenues of research:

Edema Kinetics:

Although midline shift and neurologic deterioration usually occur beyond 24–48 hours after stroke onset, edema as a response to injury begins within the first few hours. We have now obtained CSF volumetrics at serial time points after stroke from a cohort of 738 patients with strokes of varying severities. This allowed us to map the early trajectory of edema and better understand the kinetics of this process [25]. Our modeling of CSF over time in this large cohort confirmed that the majority of displacement occurs within 24 hours of stroke onset and differs significantly between those destined for edema and midline shift and those with smaller strokes and even large strokes without midline shift. The dynamics of CSF displacement may also differ by brain compartment; for example, seen earlier from the ipsilateral sulci and affecting the ventricles somewhat later. It was interesting in our preliminary data that there was also volume loss in the contralateral sulci early after stroke, suggesting that edema may be a global and not only focal hemispheric process. We intent to better evaluate compartmental CSF shifts in future studies.

Prediction of Edema:

CSF volumetrics may also facilitate early prediction of patients with a malignant trajectory. Given that a significant proportion of ΔCSF occurs within 24-hours and that this early change is reflective of subsequent midline shift and deterioration, we analyzed the accuracy and thresholds for ΔCSF to predict malignant edema in this larger cohort. We found that the risk of edema with midline shift almost doubled for every ten percent increase in ΔCSF observed at 24-hours, even adjusting for baseline measures such as age, NIHSS, and serum glucose [25]. We also demonstrated that ΔCSF is predictive of poor outcome after stroke.

Future Directions

Prospective studies will need to confirm whether CSF volumetrics can accurately and preemptively identify those stroke patients destined for deterioration and aid in triaging patients for closer monitoring and aggressive interventions (such as emerging therapies to reduce edema) [26, 27]. Studies should also compare edema biomarkers (such as ΔCSF and NWU) to determine which are most relevant to important edema outcomes and which are scalable and can be implemented most widely. For example, both these metrics are limited when evaluating edema in stroke patients who develop hemorrhagic transformation. NWU, as it employs CT-density, will provide falsely low values in the presence of high-density blood within an infarct region. Although ΔCSF can be performed in the presence of hemorrhage, the net displacement measured will reflect a combination of mass effect from stroke-related swelling and any parenchymal hematoma. We are now endeavoring to segment and quantify hemorrhage in order to disentangle these two important stroke-related complications [28].

Genetics of Cerebral Edema

In order to identify new targets for edema therapeutics, we need to better understand the critical molecular pathways responsible for propagating edema. We aim to employ a reverse translational approach to this discovery, i.e. by leveraging observed human variability in ΔCSF (accounting for stroke severity and baseline variables such as age and CSF volume), we can identify genetic variants associated with cerebral edema. The application of a quantitative biomarker of edema increases the power of this genetic discovery [29]. This translational approach has yielded novel insights in a number of diseases [30, 31]. However, it requires measuring the phenotype in thousands of subjects and pairing this to genomic data. We have now acquired both the means to measure ΔCSF in large cohorts of patients and the cohort of patients with paired imaging and genomic data from the National Institutes of Neurological Disorders and Stroke-funded GENISIS (Genetics of Early Neurological InStability after Ischemic Stroke) study. Analyzing variability in ΔCSF across this cohort will allow us to evaluate the genetic and biologic basis of cerebral edema for the first time.

Abbreviations:

CSF

cerebrospinal fluid

CT

computed tomography

FU

follow-up

MLS

midline shift

MRI

magnetic resonance imaging

NIHSS

National Institutes of Health Stroke Scale

XNAT

The Extensible Neuroimaging Archive Toolkit