3D Printing a Hyperportable, Low-cost Stroke Detection Unit
LaGuardia Studio Collaborates with Researchers to Make Diagnosing a Stroke Mobile and Affordable
In medical emergencies, time is of the essence. This is especially true of strokes, a neurologic condition brought about by the sudden loss of blood flow (ischemic) or, less often, due to the rupture of brain vessels that results in bleeding within or around the brain (hemorrhagic).1 Diagnosis of a stroke is critical to guide first-line therapy, which varies significantly depending upon the nature of the stroke. It's challenging, however, to confirm which type of stroke has occurred. Differentiating between the two requires advanced medical imaging with CT or MRI, both of which are costly, large, and generally immobile technologies seldom encountered except in hospitals and advanced imaging facilities. Consequently, many patients cannot get a confirmatory diagnosis until arrival at a clinical setting, by which time the window for administering certain brain- (and potentially life-) saving therapies has closed.
In the US alone, ~795,000 individuals are affected every year, with 1 in 4 people suffering a stroke in their lifetime. The resulting economic burden is staggering, with 34% of global total healthcare expenditure attributed to stroke, and greater than $56 billion in associated costs in the US alone. Considering the costs and limited access to imaging technology, as well as the demand for specialized technical operation and interpretation, it comes as little surprise that negative outcomes are so highly concentrated outside major global metropolitan centers and prove especially pervasive in low and middle-income countries (LMIC), perpetuating shocking imbalances in global health outcomes.
Two NYU scientists, Leeor Alon, PhD (Assistant Professor of Radiology), and Seena Dehkharghani, MD FAHA (Professor of Radiology and Neurology), both of the NYU Grossman School of Medicine's Department of Radiology, refused to accept this and began pondering how they could make stroke detection more deployable, more accessible, and more affordable. The undertaking was done in partnership with NYU's Laguardia Studio and uses two rapidly evolving technologies—3D printing/additive manufacturing and artificial intelligence/machine learning—built upon the backbone of fundamental electromagnetic principles to craft a comparatively small and lightweight, 3D printable, hyperportable stroke detection system.
Strokes and Stroke Detection
Both ischemic and hemorrhagic strokes, though indistinguishable in their immediate effects, demand radically different treatments. A hemorrhagic stroke requires interventions to control bleeding, while an ischemic stroke necessitates quick action to restore blood flow. Since about 1.9 million neurons are lost each minute in an ischemic stroke, it's crucial to diagnose and treat within the optimal 4.5-hour window. Treatment to clear blocked vessels must commence quickly once imaging rules out a hemorrhage, but the reality is that even with advanced medical centers, delays in obtaining CT or MRI scans often result in missed opportunities for treatment.
Consequently, only about 10% of ischemic stroke patients get first-line intravenous therapy; the rest receive palliative care. This results in roughly 65% of these patients dying or suffering moderate to severe disability, with only 10% making a full recovery. The prognosis for hemorrhagic stroke, particularly with ruptured aneurysms, is similarly bleak, with mortality rates topping 40% within 30 days in some studies.
Despite their importance in neurological care, CT and MRI have significant challenges including high costs, large size, and complex operation and interpretation needs. CT scans, which use high-energy X-rays, pose risks of ionization potentially leading to cancer or tissue damage. MRIs, though free from ionization risks, can cause tissue heating and are dangerous around incompatible metals. Screening for MRI compatibility can delay treatment, which is problematic during stroke emergencies. Additionally, MRI scans require lengthy protocols, extending the time needed for initial assessment.
Alon and Dehkharghani saw large and untapped possibilities to liberate medical imaging from its conventional constraints and to achieve a precise, information-rich, and even directly diagnostic scan with reduced overhead and dependence on sophisticated scanner operation and image interpretation. But how?
"How do we make the technology to do this more readily available," asks Alon, "especially in places where there may be only a single radiologist for an entire country?"
Dielectrography and Design
"I have this idea," Alon said to Dehkharghani after the two met at a seminar presented by Dehkharghani. "You have this electromagnetic wave, this microwave like your cell phone. What if we transmit it through the body? Could we detect changes in the tissue that occur with stroke? For example, if you have blood in the brain, could we measure the way that it defracts the waves?"
The idea was intriguing, but the use of microwaves in medical imaging technology has generally stagnated and has largely been ignored by comparison to radiofrequency, X-ray, gamma ray, and ultrasonic techniques. While they can be used at low operating power, their interactions with matter are challenging to localize and seem at first poorly suited for imaging of the sort employed in other medical imaging. The two began brainstorming. Both ischemic and hemorrhagic strokes result in known alterations in electrical properties. So why stop at image formation, they asked one another, only partly in jest.
“There’s no rule saying we have to create an image out of a physiologic, anatomic, or pathologic probe for it to be useful or informative,” Dehkharghani asserts, acknowledging the irony of such an assertion from a radiologist. The information content of the proposed microwave interactions can be extremely rich, in particular if a system could leverage contemporary antenna technology to achieve wide bands of interrogation across the microwave spectrum from a single antenna. They reasoned that they could then distribute an array of such antennas around the head to increase its performance and sensitivity to disease.
They recognized the challenge of processing the extensive data for traditional algorithms but identified deep learning as a promising tool to decipher disease markers like those of strokes. If such a neural network could be trained to learn specific signatures of diseases (e.g. hemorrhagic or ischemic stroke, among others) from the vast ocean of electromagnetic interactions prescribed in their hypothetical system. They admired their proposed system’s dual benefits of affordability and compactness, paired with high-speed performance. Their concept involved an ultra-wideband microwave array encircling a patient’s head, which they termed dielectrography, for its precise detection of tissue dielectric properties.
Following promising initial computer simulations, Alon and Dehkharghani were faced with the questions of how to ensure low cost, consistent performance, and reproducibility. In hopes of expediting the development and production process, they reached out to the NYU LaGuardia Studio for their expertise in the field of 3D content development and additive manufacturing (AM).
The optimal design would consider the necessary dimensions of the helmet for human use while accommodating the requisite electronic circuitry and antennas; however, any design that would allow the helmet to fit comfortably and securely over the head would be larger than the printing bed of available 3D printers. Because cost was a primary consideration, requiring a very large 3D printer that could print the helmet as a single piece would undermine their primary goals. Vito Ciancia, an AI 3D media engineer at the Studio, took on the challenge of developing and printing a device that would not currently fit in any print bed that was currently available.
Ciancia came up with the idea for a version that could be printed in modular sections that could then be assembled (without the use of magnets, which interfere with the scan) along with the required electronics and cabling that would power the system itself. Working in close collaboration with Alon and Dehkharghani over several months to refine the idea, Ciancia designed the geometry for this new iteration.
With a solution for the helmet in place, they moved next to the question: how best to position an acutely ill or traumatized patient quickly and with minimal distress. Ciancia and the researchers sought a design that integrated the entire unit, from helmet to patient support, into a cohesive whole. They attached the helmet to a mobile gurney upon which a patient could lie and to which the helmet could be mounted. Using LaGuardia Studio's 3D scanning service, members of the team captured Ciancia's head and torso. In addition to the new scans, a section of the mobile gurney was modeled. The three 3D modeled components—consisting helmet, head/torso, and gurney—were used as constraints to design and validate the mounting, ergonomic patient positioning, and leveling system.
The final helmet was substantially smaller and lighter—essentially transportable by hand—and therefore far more mobile than any existing technology. Alon and Dehkharghani tested their new device, and the results were a resounding success. The unit fulfilled every projection and aim from the initial conception of the project.
"The properties that we're interested in capturing," says Dehkharghani, "are the disturbances and perturbations in dielectric properties of tissues that occur during a stroke or in other neurological disease. The changes of interest to us at the moment would be relatively localized or locally distributed in a brain region, like ischemic or hemorrhagic strokes. We’re extremely sensitive to the dielectric changes that are engendered by the presence of a hemorrhage, for example, and we can detect it in real-time using an incomparably low-cost, mobile, and low-power instrument."
Dehkharghani sees the detection unit as being a potentially vital addition to the imaging and diagnostic armament in a wide number of situations. "An ambulance is an obvious one that has already been used to house mobile CT units, but we’re intent on pushing the technology far upstream to pre-clinical environments like nursing homes, sporting events, shopping malls, mass casualty scenarios, and so forth. Any environment for which we’d desire the ability to quickly deploy technology for the detection of hemorrhage or ischemic stroke. Looking further down the line, we expect to be very sensitive to overall water content, which means that we should be able to train the system to detect hydrocephalus, for example, which can be a life-threatening consequence of several childhood conditions, again disproportionately affecting LMIC."
Their next consideration was improving the interpretation of the scans so that an accurate initial diagnosis could be made and appropriate management initiated. Here again, Dehkharghani and Alon point to the shocking imbalances in global access, with large swaths of the global population lacking access to advanced radiologic interpretation. For such a technique to reach its full potential, it could not be left dependent on expert interpretation where none exists, spurring their motivation to circumvent the need for image formation and image interpretation in favor of intelligent, precision diagnostic instruments. For this, the duo turned to artificial intelligence and machine learning, which had recently proven effective in detecting neurologic disease from other modalities.2
AI in medical diagnosis is a tool meant to augment and assist the skills of a human professional. In cases where a radiologist may not be readily available, such as in an emergency or a remote location, machine learning algorithms and artificial neural networks can assist doctors and first responders with rapid and accurate results like hemorrhage detection or exclusion. It can quickly analyze images, significantly reducing the time it takes for a radiologist to review the scans. This speed can be critical during emergencies and help streamline the diagnostic process. It can also aid radiologists in detecting subtle abnormalities that might otherwise be missed. Machine learning models can also be trained on large datasets to identify patterns and indicators of specific conditions, leading to increasingly accurate diagnoses.
Despite their initial emphasis on direct-to-diagnosis technology, their team has returned more recently to the elusive task of anatomic image reconstruction from microwave reflections. Using other dedicated neural network architecture, the team feels that they have solved one of the most recalcitrant and long-standing challenges in their field, having already achieved the production of anatomically realistic, MRI-like tomographic brain images using their ultra-wideband microwave helmet.
Dehkharghani says the quality of scanning achieved by their unit is well-suited for making use of emerging AI-assisted diagnostics: "The richness of the data spans wide ranges of interrogation for the physical properties of interest, in fact much more so than in MRI, ultrasound, or CT, where the spectrum is comparatively quite narrow. We're taking advantage of these properties and some of the advances in these designs to interrogate a huge aperture around the spectrum of interest. That richness is really well-suited to taking advantage of machine learning algorithms. A very big part of what we have in mind is taking advantage of the advancements in machine learning."
He adds, “We’re most excited for having achieved our aims while remaining within our initial projections and limits for safety and size.”
He foresees the potential for a scanning unit that captures a sufficiently-detailed collection of data and is informed by a sufficiently-robust data library to make accurate diagnosis without the need for the image itself.
"Most medical imaging exists so that a specialist can review it, render an interpretation, and inform a management plan for a patient. But there's no rule saying that the actual image is critical to that process. It's just been practice and convention, because well, what would be the alternative? You can't expect the machine or interpreter to offer a diagnosis without the image, can you? We'd like to flip that script and show that perhaps the machine can tell you what's wrong if the input is sufficiently rich."
The theoretical impact of any new technology must be tempered by the reality of a situation, but Alon and Dehkharghani are excited about their unit's potential to be a transformative addition to emergency medicine, especially as both additive manufacturing and artificial intelligence continue to expand and improve. They have already published their initial computer simulation results, and have presented results of their physical system on the international stage at major bioscientific and biomedical society proceedings while preparing subsequent manuscripts. They have applied to NYU Langone's Institutional Review Board for permission to begin using the latest version of the detection unit on stroke patients as a way to gather additional data about its efficacy.
Both are excited about the potential of the unit to make a far-reaching impact on healthcare, especially in those areas still unmet by existing imaging technology.
Dehkharghani concludes, "Stroke is on the rise, and while we’ve made tremendous progress in diagnosing and managing it, half of even optimally treated patients are dead or dependent at 90 days. The overall prevalence is projected to increase by more than 20% by 2030 by comparison to 2012. As has always been the case, while strokes affect people of all ages and walks of life, the burden is heavily and disproportionately on the socioeconomically disadvantaged. We aren’t here to evoke fear but to invoke awareness. Something needs to be done; in fact, many things need to be done, but we hope to address at least one with this technology. More so than anything, we’re in pursuit of the democratization of access."
- Stroke Facts (Center for Disease Control and Prevention)
- Zhang, B., Shi, H., & Wang, H. (2023). Machine Learning and AI in Cancer Prognosis, Prediction, and Treatment Selection: A Critical Approach. Journal of Multidisciplinary Healthcare, 16, 1779-1791.