By Keith Loria
Vol. 21 No. 8 P. 14
AI helps detect stroke and other neurovascular brain abnormalities.
Across radiology, deep learning technology has fueled a new generation of startups focused on leveraging AI to improve imaging diagnosis. Among these startups, AI-enabled stroke triage is perhaps the most popular application.
“This is, in large part, driven by the widespread incidence of stroke and neurovascular disease across the US and the world, as well as the fact that imaging plays a key role in guiding therapeutic decision making,” says Peter Chang, MD, a radiologist and cofounder of Avicenna.AI, a medical imaging software company providing solutions for emergency radiology.
Additionally, several key steps of the stroke triage pipeline have been supported by software analytics for many years and, thus, provide a natural point of entry for the next generation of AI-enabled tools.
Pooja Rao, PhD, cofounder and head of research and development at Qure.ai, notes that the industry is not quite at the stage where diagnostic radiologists rely on AI tools completely, though, and, in many centers, AI is being primarily used for research and development. “In contrast, the adoption of AI imaging tools to alert for strokes and other neurovascular abnormalities is somewhat higher among other specialties, such as interventional radiology and neurosurgery,” she says.
This is important because, in patients having a stroke, the blood supply to the brain is affected and 2 million brain cells—neurons—die every minute due to a lack of blood supply. “Treating strokes and traumatic brain injuries is an extremely complex process from onset to diagnosis and then treatment—every step requires the attention of highly trained professionals,” Rao says. “These professionals are in short supply and typically overburdened. In this scenario, every bit of help that AI tools can provide directly improves patient outcomes and reduces the chances of death or permanent disability.”
Greg Albers, MD, director of the Stanford Stroke Center in California, notes that, as medical professionals continue to care for patients during and after the pandemic, the “new normal” will include even more technology-forward practices, and AI will be a big part of that. “AI is going crazy right now. I think there are about 15 hemorrhage detection products that have gone through the FDA, at least from a notification perspective,” Albers says. “If every radiology department in the country has their own detection module, how will that translate into practice? Will people wind up using their own or will there eventually be a small number of companies that dominate the market?”
The answer to that question will be the key as to whether AI, in these types of cases, becomes a success. The important factor to keep in mind is that AI is not the all-powerful solution to all things that many may think it is.
“The current AI products are really there to flag things for physicians to take a look at, and that’s a bit of a disconnect between what many people think AI is supposed to be doing and what it actually does,” Albers says. “At this point, it is there to help the physician see things that might have been missed, but it’s not there to take over the role of the radiologist.”
In that regard, Albers helped launch RapidAI, a cerebrovascular imaging platform that empowers clinicians to make faster, more accurate diagnostic and treatment decisions for stroke patients. “There are a number of people reading scans in the midst of an emergency situation. My area is stroke, and stroke treatment happens very quickly at all hours of the day or night,” he says. “You [often] don’t have the luxury of getting a neuroradiologist to look at scans. So, when you have less experienced people looking at scans, having an AI helper flagging you to things can be particularly helpful.” There are more than 1,500 customers in 50 countries using the RapidAI tool, mostly expert stroke centers and primary stroke centers.
Gal Yaniv, director of endovascular neurosurgery at Sheba Tel HaShomer Hospital in Israel and cofounder of Aidoc, notes that the hospital has been utilizing AI to help diagnose time-sensitive pathologies such as brain bleeds or large vessel occlusions. “We use this algorithm mostly in the [emergency department] setting, but other departments are using it to make sure we didn’t miss anything while reading scans,” Yaniv says. “We can prioritize scans in urgent need of review and also expedite the treatment of urgent patients through the workflow.”
Knowledge Is Power
Avicenna.AI’s software solutions leverage deep learning to improve the speed and accuracy of stroke diagnosis. Chang explains that following the step-wise progression of stroke diagnosis in a hospital, the AI system is able to exclude the presence of hemorrhage in the brain, eg, patients with hemorrhagic stroke receive completely different treatment than those without a bleed; identify the presence of a clot in critical areas of the brain that are amenable to surgical intervention; and quantify the amount of infarcted tissue in the brain, to ensure that patients designated for intervention will benefit from treatment.
“There are two key benefits of AI-enabled diagnosis. First, in busy emergency [departments] and stroke centers, our deep learning system is able to rapidly identify the key markers of stroke and alert a physician when a specific patient may need to be urgently evaluated,” Chang says. “Second, in smaller community hospitals without 24-hour access to stroke experts, the AI system can be used to increase the confidence and objectivity of interpretation and, if needed, identify patients whose care needs to be escalated to a dedicated stroke center.”
Qure.ai’s head CT scan tool, qER, which recently received FDA 510(k) clearance, detects the presence of bleeds, cranial fractures, a mass effect, or midline shifts on plain CT scans. “As soon as one of these critical abnormalities is detected, a clinician is notified,” Rao says. “The tool generates a notification much before a radiologist might have otherwise seen the scan, so it can get immediate attention. Our studies show a potential 97% decrease in time to action when qER is deployed.”
“Our target turnaround time for emergent studies is 30 minutes, but leveraging AI for acutely critical conditions enables us to shorten that time,” says Benjamin W. Strong, MD, chief medical officer of vRad. “For conditions such as intracranial hemorrhage, time is of the essence and those precious minutes can be life changing for our patients. We have done extensive validation of the qER solution, and we are excited to continue partnering with Qure.ai and improving care for our patients.”
Tools in Practice
Yaniv says radiologists today are under a great deal of pressure with voluminous scans to read and consultations; this can sometimes lead to missing a brain bleed or fracture. The Aidoc algorithm helps them avoid misses. It also helps identify patients who need to be seen more urgently.
Daniel S. Chow, MD, assistant professor-in-residence in the department of radiological sciences and codirector of the Center for Artificial Intelligence in Diagnostic Medicine for the University of California, Irvine, has been utilizing Avicenna.AI’s software in a research setting for more than six months. “Having that real-time alert is helpful,” Chow says. “One of the things we are seeing is that medical centers and providers are increasingly expected to provide quality care and to do so more efficiently. That can translate into longer times to get to those studies that have critical findings. So, one of the nice things it does is provide an alert to say a patient may have a hemorrhage so that exam needs to be evaluated first.”
Still, it’s important to remember that it’s not the AI technology that’s the game changer in these instances but the case-based applications. “One of the things I like about Avicenna is that it’s looking at the critical needs and critical challenges and working to solve the challenges with AI,” Chow says.
In stroke cases, the AI tools being used are hemorrhage detection, large vessel occlusion (LVO) detection, and aspects. “The LVO can be very useful because these are patients that need to have a thrombectomy and need that clot removed, and you want to do that as quickly as possible,” Albers says. “At Stanford, we have multiple hospitals who send patients to us and our hospitals use Rapid as our AI program. What we can do is see a patient [who has an LVO] at one of our outside hospitals and find out about it [at our center] at the same time, so we can be prepared to accept the patient in transfer and bring them over quickly.” Without that capability, it may have taken an hour or two for a scan to be read; this AI product does it in a couple of minutes.
Rao notes Qure.ai’s triage and notification tool is utilized by large health care providers or teleradiology groups where the volumes of scans are high. “They use it to prioritize critical head CT scans in their radiology worklists—CTs that need urgent attention are flagged even before they are opened for viewing,” she says. “The manual process of opening and viewing each CT to decide which should be examined first is handled by our algorithm, saving precious time for the radiologist. As a result, emergency department physicians or other clinicians requesting time-sensitive scans have shorter wait times.”
Although Avicenna.AI’s initial focus relates to stroke imaging, the same deep learning technology can be, and has been, applied successfully to many other disease processes from head to toe.
“High performing deep learning algorithms require tremendous volume and diversity of data for optimal performance,” Chang says. “Our AI solutions are trained using data from hundreds of hospitals across the US, Europe, and Asia, with partners ranging from large academic medical centers to teleradiology groups and community hospitals.”
“It’s well known in the field of AI that the most accurate algorithms are those that are trained on the largest and most diverse data sets,” Rao says. “That is what we strive for—our chest X-ray product qXR, for instance, is trained using more than 3 million X-rays. We get our training data through our research collaborations with universities and hospitals around the world.”
While Qure.ai’s algorithms can currently detect, localize, and quantify a number of abnormalities in chest X-rays and head CT scans, Rao notes the next step, which isn’t far in the future, is the capability for its AI tools to draft a full chest X-ray or a CT scan report.
The data for the Aidoc tool came from a diverse number of hospitals with a range of data from different scanners. “We needed to train the computer for what it’s looking for, and we specifically focused on the hardest cases,” Rao says. “As a clinician, one of the most annoying things is to get a lot of false-positives. But, because of the wide range of scans, we are not seeing a lot of those.”
The algorithms for RapidAI are created by feeding data into the neural network, allowing the computer to learn. For brain hemorrhages, for example, 1,000 cases with hemorrhages circled and 1,000 cases with no hemorrhages circled may be used, and the computer will learn what a hemorrhage looks like.
“You train and train and train until the model is very good,” Albers says. “Then you do some pilot testing and provide a new batch of hemorrhages. You then see what the sensitivity is and, hopefully, it’s going to be in the range of 95% to make it a useful tool.”
It’s equally important to check for false-positives and use different scanners from different hospitals to obtain the data, he adds. Albers also reminds users that these tools are not going to replace physicians; they are designed to help radiologists do their jobs better.
“There can be huge benefits to all this,” Albers says. “The sooner you treat the patient, the better, which is a compelling reason to have AI software available to help you out.”
— Keith Loria is a freelance writer based in Oakton, Virginia. He is a frequent contributor to Radiology Today.