Dr. Rajat Dhar, left, and Dr. Saurav Das
Dr. Rajat Dhar, left, and Dr. Saurav Das

A conversation with Dr. Rajat Dhar, MD, Associate Professor of Neurology and Neuro-critical care, Washington University School of Medicine, St. Louis, MO.

Interviewed by Dr. Saurav Das, MD, Fellow in Vascular Neurology, Washington University School of Medicine, St. Louis, MO.

They will be discussing the article “Automated Quantification of Reduced Sulcal Volume Identifies Early Brain Injury After Aneurysmal Subarachnoid Hemorrhage,” published in Stroke.

Dr. Das: Dr. Dhar, on behalf of the Blogging Stroke team, we welcome you to this author interview.  I read with great interest your paper pertaining to the automated estimation of selective sulcal volume (SSV) to quantify global cerebral edema (GCE) from early brain injury (EBI) in aneurysmal subarachnoid hemorrhage (aSAH). This is an important paper as our understanding of clinical outcomes following aSAH is shifting from vasospasm induced delayed cerebral ischemia (DCI) towards GCE from EBI. Also, we currently do not have the tools to measure GCE accurately.

This research uses a “deep learning-based approach” for the analysis of serial CT scans to measure SSV. Many of our readers may not be familiar with the use of artificial intelligence (AI) in image analysis. I will begin by requesting you to explain what deep learning is.

Dr. Dhar: Applications of artificial intelligence, specifically machine learning, to the realm of biomedical image analysis have been growing exponentially over the past few years. AI is well-suited to image analysis because, at its core, machine learning seeks to find patterns in data, and images are just patterns of intensity and location data. Machine learning algorithms can be trained to learn from labeled data. For example, to determine what regions of a scan represent blood vs. brain vs. CSF is called a segmentation task. We can use machine learning to perform a segmentation task on new imaging data. AI algorithms can perform image analysis in a fast and reproducible way, eliminating the need for time-intensive human input. They can measure volumes of similar brain structures over serial time points more objectively and accurately than one or more humans may be able to.

There are various forms of machine learning approaches that may be applied to image analysis, but recently neural network architectures have demonstrated superiority for performing complex classification and segmentation tasks. Deep learning is then, in essence, the type of machine learning that uses neural networks to learn and make predictions. A major advantage of deep learning over traditional AI for image analysis is that it can extract features within its many hidden layers, without human planning or input. This is in contrast to older machine learning approaches that typically require preformulated features, selected by humans. For example, in this study, we used a well-accepted architecture for image segmentation, the U-Net, so-named because it has contracting (encoding) and expanding (decoding) pathways. Encoding allows the image to be transformed into many feature representations while the decoder projects discriminative features back onto the original image to provide a probabilistic segmentation map. Such deep learning approaches have been shown to perform very well in segmenting brains and many other image types into relevant parts. The downside of deep learning networks is that its output is typically “black box” — meaning you do not always know what features were used and how it came to its conclusion. For image segmentation, a practical not philosophical task, this is not much of a concern, but for medical decision-making, it can be. Further, I find it amusing to think that we are using neural networks, modeled on the architecture of brain connections, to visualize and understand the brain itself! Truly meta-neural networks!

Dr. Das: In this study, 66% of patients developed hydrocephalus and required external ventricular drain (EVD) placement. In these patients, other physiological parameters like intracranial pressure, cerebral perfusion pressure measurements, etc., may be available. Intuitively, SSV may be a function of these physiological parameters, rate, and duration of cerebrospinal fluid (CSF) drainage, etc. What are your thoughts? 

Dr. Dhar: In our study, we modeled global cerebral edema as a consequence of early brain injury using the metric, SSV (the volume of sulci remaining in the brain above the level of the ventricle). Many of these SAH patients did have EVDs that allowed both CSF diversion and ICP monitoring. For one, this allowed us to correlate imaging with elevations in ICP (which we included in the determination of “symptomatic edema”). We demonstrated that those with either deterioration from edema or high ICP had much lower SSV. Secondly, you are correct that CSF drainage could influence CSF volume. However, such drainage would largely reduce CSF volume from the ventricular system. We did observe this. However, it should not directly influence SSV, which only measures sulcal volume on the surface of the brain. Nonetheless, it is possible that drainage of CSF to treat severe hydrocephalus (common early after SAH) could lead to a reduction in global edema. For this reason, we analyzed the SSV from scans before an EVD was placed compared with those in the 72 hours after placement in 100 SAH patients from our cohort. We found no overall change in SSV in such patients, suggesting that CSF drainage does not significantly affect SSV. We did confirm a significant reduction in ventricular volume (unpublished), as would be expected with CSF drainage. As you state, it is interesting to wonder how cerebral perfusion pressure and other physiologic variables could influence edema formation. We did not collect such data in this study, but now that we have a quantitative dynamic biomarker of edema, we plan to evaluate the effect of such physiologic modulators on edema formation in future SAH studies.

Dr. Das: It is interesting that early SSV has a stronger effect in younger patients and remains predictive of GCE only in individuals less than 70 years of age. I believe this is due to cerebral atrophy with aging. Do we have any data regarding sulcal characteristics including SSV change with aging in the normal population?

Dr. Dhar: Yes, I agree, it seems that aging and associated sulcal atrophy are protective against the deleterious consequences of edema formation. The same phenomenon has been noted by our group and others in ischemic stroke, where this “intracranial reserve” (CSF volume as a proportion of cranial volume) has been shown to be a protective factor against malignant edema. Although, conversely, it is a risk factor for impaired neurologic recovery. So, atrophy is a double-edged sword in stroke. Here in SAH, we applied SSV as a quantitative marker of edema to illustrate that younger patients are similarly more at risk for edema-related poor outcomes. Yet, we know that aging is a risk factor for poor outcomes after SAH, independent of edema. So, it may be that the deleterious neuropsychological aspects of aging and atrophy counteract any protective signal against edema as SAH patients get older.

In our stroke studies, we have analyzed early admission CTs (i.e., before signs of stroke become visible) to estimate baseline CSF volume and atrophy in relation to age in almost one thousand stroke patients. We did show a very strong correlation of aging with atrophy, as measured by higher total CSF volumes and more intracranial reserve. We do not have such normative data on SSV (sulcal volume) specifically, but we did attempt to model it within our SAH cohort — focusing on those without edema and evaluating only baseline CTs. We showed a similar but more noisy relationship (figure shown in our Supplemental Data). We plan on studying this more: In fact, one hypothesis is that the admission SSV of an individual SAH patient can be “normalized” to expected age-related sulcal volume using such data. This would allow us not only to rely on “absolute SSV” (whether it is less than 5-ml, a cutoff we demonstrated useful in this study), but to extend this analysis to observed vs. expected SSV — which may serve as a more sensitive age-specific marker of edema right on arrival to the ED and as edema is developing.

Dr. Das: Your team has published a recent paper in Stroke regarding the use of CSF volumetrics in predicting cerebral edema and midline shift following malignant strokes. What are the conceptual similarities and differences in the use of AI in quantifying cerebral edema in these two different pathologies?

Dr. Dhar: We actually started with the same U-net based CSF segmentation algorithm for both studies. However, we had to validate it in the SAH population against manual results to ensure it was reliable in the presence of blood (which it was). Additionally, for our stroke studies, we had focused on global and recently hemispheric changes in CSF volume, not SSV (which is really particular to SAH and more diffuse brain injuries). So, we had to develop algorithms that separated total brain CSF into ventricular and sulcal compartments, and then we divided the sulcal compartment into SSV (i.e., that located above the ventricle). These extra steps were unique to the SAH project. These illustrate how SAH induces more diffuse edema resulting in effacement of bilateral brain sulci especially near the vertex, as opposed to ischemic stroke that induces ipsilateral hemispheric reduction in CSF from both sulci and ventricles. Nonetheless, the overriding concept that unifies these two projects is that we can learn more about edema by quantifying it and modeling how it changes over time. Just so you know, we have also worked on automatically quantifying perihematomal edema after intraparenchymal hemorrhage using related imaging algorithms and published the results of that algorithm recently also in this journal. The bigger pursuit is to identify how clinical and biological factors modulate critical responses of the brain to different types of injury. Accurate measurement of progression and severity of edema will help this scientific pursuit. We are currently involved in genetics studies, for example, in all three disease states, that will allow us to evaluate which genetic markers are related with enhanced vs. protective of edema formation. We hope this research will lead to a better understanding of edema in all types of brain injuries, and the development of new targeted therapies.

Dr. Das: Are these AI tools available to physicians at other centers to calculate the sulcal volume and quantify GCE in the form of web-based or mobile phone-based applications? If not, how do you plan to disseminate such tools for your results to be used at other centers?

Dr. Dhar: This is an excellent question and the subject of much active work in our lab. Once we validate such algorithms and imaging biomarkers (as we have begun doing in these publications), then we next need to figure out how best to deploy tools that both clinicians and researchers can use. That means web- or cloud-based applications that can analyze scans both in large numbers (for research studies) and rapidly at the bedside (to facilitate precision medicine for individual patients). We are certainly working on that, albeit with less of a footprint than large AI-focused companies who have made great strides in developing mobile-ready applications! Our initial focus has been on accelerating “big data” edema research through automated imaging algorithms, with a next goal to move to clinical applications as the algorithms are further tested and refined for use at the bedside. This will require rigorous reliability testing.

To that first end, we have created a novel stroke imaging repository to centralize, archive, process, and analyze images from all around the world: This is based on an imaging platform called XNAT, developed by our collaborator here at Washington University in St. Louis, Dr. Daniel Marcus. We have created a specialized stroke imaging platform called SNIPR: the Stroke NeuroImaging Phenotype Repository. It currently houses several thousand serial scans from several sites and stroke studies. Importantly, what SNIPR allows is execution of image processing modules such as CSF segmentation and measurement of edema — from hundreds of scans, processed on parallel computing clusters at high speeds. We are now finalizing these pipelines and then hope to invite collaborators to carry out their imaging research using this platform. We also hope to develop web-based tools and eventually mobile phone-based applications, but are focusing on SNIPR for now.

Dr. Das: The first author of this paper, Jane Yuan, is your mentee and now a third-year medical student at Washington University in St. Louis. I will end by asking if you have any message for trainees and students interested in stroke-related AI research.

Dr. Dhar: I think we actually had four trainees involved in this particular project, and I have noticed greater student and trainee interest in the kind of research that we are doing, over the past few years. I have several students in the lab working on various AI-related projects, including other topics like radiomics and natural language processing. I think students really get excited about how AI research can accelerate and transform traditional research paradigms and provide novel approaches to dissect diseases that they can be intimately involved in. Jane is an outstanding example. We started with a fundamental question: Can we measure edema after SAH more accurately than existing scales and scores? She was able to work on this from start to finish, which I hope was satisfying and contributed to a breadth of learning.

In fact, AI research is great because trainees can help at so many levels; training a computer really requires a village (of trainees and experts, both). You need capable co-investigators to review quality of images, label regions for supervised AI learning, and then evaluate the results critically. You can also learn how to analyze large datasets with quantitative measurements at several time points, building statistical and data management skills. We have an excellent team to support such work, including our co-investigators Drs. Yasheng Chen and Atul Kumar, who are biomedical/computer engineers who specialize in imaging research. They worked hand-in-hand with the students and me to develop, refine, and test these algorithms. The students often offer insightful comments and notice things in review of CTs that we did not. It was Jane who recognized the age-dependency of SSV and suggested building a regression model that incorporated an interaction term of age and SSV. Her finding really offered novel insights into how edema affects SAH patients. Ultimately, AI is coming (in fact, it’s already here!), and getting involved in such research allows trainees to gain skills and insights into how this kind of work will affect medicine in decades to come. They are the next generation, so it’s very rewarding to have them involved in a research paradigm that I believe will be an integral part of their scientific landscape.

Dr. Das: Dr. Dhar, it was a pleasure having you with us, and thanks for your contribution.

Disclosure: Dr. Dhar and Dr. Das hold affiliation to the same institution, Washington University School of Medicine, St. Louis, MO.