Company Spotlight: Methinks CTA Stroke - Triage AI for Faster Interventions

Maya: In stroke care, saving three minutes can mean saving millions of neurons. Today, we are looking at a new AI tool that flags severe strokes on brain scans and notifies specialists in an average of 3.30 minutes.

Sam: Just a quick note before we start: Maya and I are AI-generated voices. Today we are summarizing the official FDA 510(k) summary for Methinks CTA Stroke, published by the U.S. Food and Drug Administration on August 20, 2025.

Maya: This is a Company Spotlight episode, but we are not reading a corporate brochure. This is the public regulatory document clearing the device for the US market, which means we get to see exactly what the software does, the real testing numbers, and its limitations.

Sam: Exactly. And for hospital networks buying stroke AI, this is the highest-stakes procurement category there is. So Maya, what exactly is Methinks CTA Stroke?

Maya: According to the summary, it is a software-only device classified as a Radiological Computer Aided Triage And Notification Software. It analyzes Computed Tomography Angiography, or CTA, images of the brain in the acute setting to identify suspected Large Vessel Occlusions, which they abbreviate as LVO.

Sam: And an LVO is exactly what it sounds like, a major blockage in one of the main arteries feeding the brain. But the document mentions this is a parallel workflow tool. What does that mean for the doctors actually using it?

Maya: It means the AI operates alongside the normal hospital procedures, not instead of them. The document states it is used to notify an appropriate medical specialist of these findings in parallel to standard of care image interpretation. It does not remove cases from the standard review.

Sam: Right, the FDA summary repeats multiple times that the identification of suspected findings is not for diagnostic use beyond notification. It sends a notification to a neurovascular specialist recommending they review the images, which they can preview through an image viewer.

Maya: And there is a strict warning about those preview images. The summary explicitly says the previewed images are for informational purposes only. The clinicians still have to view the non-compressed images on a diagnostic viewer and evaluate the patient fully before making any care-related decisions.

Sam: Because this is a 510(k) clearance, Methinks did not have to prove their device is completely novel. They just had to prove it is substantially equivalent to a legally marketed predicate device. What was the predicate here?

Maya: The predicate is a device called ContaCT, manufactured by Viz.AI, Inc, which was previously cleared under DEN170073. The FDA summary confirms both devices have the identical intended use and similar technological characteristics.

Sam: So if they are substantially equivalent, is there any difference at all between the two?

Maya: There is one specific anatomical difference noted. Both devices analyze the terminal ICA and MCA-M1 vessels. But the Methinks subject device identifies an additional type of occlusion in the MCA-M2 vessel, which the summary notes the predicate does not do.

Sam: Okay, so Methinks covers the ICA, the MCA-M1, and adds the MCA-M2. The document says this addition does not raise different questions of safety and effectiveness. But let us look at the actual performance data. How did they test this AI?

Maya: They ran a retrospective, blinded, multicenter, multinational study. It included 336 total cases. Of those, 110 were positive for an LVO, and 226 were negative.

Sam: Retrospective is a key word there. That means they used historical data, not a live, prospective clinical trial. Who decided what the correct answer was for those 336 scans?

Maya: The ground truth was established by two US board certified neuroradiologists who read the cases. If they disagreed, a third ground truther was brought in, and the final answer was based on majority vote.

Sam: That is a standard, robust way to establish ground truth in imaging AI. The summary also notes they made sure the institutions used for the validation study were completely different from the institutions used to train the algorithm. They even checked countries, states, and ZIP codes to ensure separation.

Maya: That is critical to prove the AI actually generalizes and has not just memorized the training data. So, on to the main results. For detecting an LVO, the sensitivity was 98.2%, and the specificity was 91.6%.

Sam: Those are strong top-line numbers. Sensitivity of 98.2% means it barely misses any actual LVOs. Specificity of 91.6% means it is pretty good at ignoring scans that do not have an LVO. But the summary breaks this down into several subgroups, which is where things get really interesting.

Maya: Let us start with that speed claim I mentioned at the top. The summary lists the time to notification, which includes getting the DICOM exam, de-identifying it, analyzing it, and sending a notification back to the PACS and clients.

Sam: Right, and the mean time for all that was 3.30 minutes. The 95% confidence interval was incredibly tight, from 3.23 to 3.36 minutes.

Maya: In a hospital setting, getting a highly accurate alert about a major stroke just three and a half minutes after the scan is acquired is exactly why these tools are being bought. But the subgroup tables show the algorithm does not perform identically across all patients.

Sam: I noticed the age breakdown. For patients under 50 years old, which was only 8 positive cases and 40 negative cases, sensitivity was 100.0% and specificity was 97.5%.

Maya: But look at the older patients. For patients over 70 years old, there were 61 positive cases and 80 negative cases. Sensitivity stayed high at 98.4%, but the specificity dropped to 87.5%.

Sam: That is a noticeable drop. A specificity of 87.5% means more false alarms for the older demographic. Does the document explain why that happens?

Maya: It does not give a medical explanation, but we can look at another table that lists non-LVO subgroups to see where the algorithm gets confused. When the algorithm looked at scans with Chronic Lesions, tested on 101 cases, the specificity was 86.1%.

Sam: Ah, and older patients are more likely to have chronic lesions or messy vessels from a lifetime of wear and tear. There are a couple of other tricky categories in this non-LVO table too.

Maya: Yes. For patients with a Stenosis greater than 51% but not an LVO, there were 27 cases tested. The specificity there fell to 74.1%. And for a category labeled Other Acute Ischemic Events, with 61 cases, specificity was 78.7%.

Sam: So if a patient has severe narrowing, or stenosis, of an artery, the AI is much more likely to flag it as a complete occlusion. That makes sense. It sees a tight vessel and sounds the alarm.

Maya: Exactly. On the ward, this means the neurointerventionalist is going to get a notification on their phone, open the preview image, and realize it is a false positive caused by severe stenosis, not a fresh clot.

Sam: Which reinforces why the FDA insists these tools are just for triage, and why clinicians must do a full evaluation. What about the gender breakdown? Any disparities there?

Maya: It was fairly balanced. They tested 53 positive cases in males and 57 in females. For females, sensitivity was 100.0% and specificity was 91.0%. For males, sensitivity was 96.2% and specificity was 92.2%.

Sam: We should also talk about that new vessel they added, the MCA-M2. The table breaking down performance by LVO subgroup shows they had 25 cases of MCA-M2 occlusions. The sensitivity for that specific vessel was 92.0%.

Maya: Compare that to the other vessels. For the ICA, based on 35 cases, sensitivity was 100.0%. For the MCA-M1, with 44 cases, sensitivity was also 100.0%. So the algorithm is slightly less sensitive for the M2 segment, but 92.0% is still a solid capture rate for a vessel further out in the brain.

Sam: Something IT buyers always want to know is whether an AI tool only works on certain hospital scanners. The summary includes a breakdown by vendor machine. What did they test?

Maya: They tested across four major vendors. Siemens had the most data with 35 positive and 83 negative cases, showing 97.1% sensitivity and 92.8% specificity. Philips was also well represented with 34 positive and 82 negative cases, yielding 97.1% sensitivity and 91.5% specificity.

Sam: They also tested General Electric and Toshiba or Canon scanners, though with fewer cases. General Electric had 20 positive cases and 31 negative cases. Interestingly, sensitivity for GE was 100.0%, but specificity was lower at 83.9%.

Maya: And for Toshiba or Canon, with 21 positive and 30 negative cases, sensitivity was 100.0% and specificity was 96.7%. So it is clearly vendor-agnostic, though performance fluctuates slightly depending on the machine.

Sam: Slice thickness of the scan also seems to matter. For scans with a slice thickness between 0.3mm and 1mm, they had 91 positive cases and a sensitivity of 97.8%. But for slightly thicker slices, between 1mm and 1.3mm, they only tested 19 positive cases, which hit 100.0% sensitivity but specificity dropped to 82.1%.

Maya: This level of detail in the 510(k) summary is exactly what hospital procurement teams need to look at. You have to know what kind of scanners your hospital uses and what slice protocols your radiology department runs to understand how this algorithm will behave in your specific environment.

Sam: There is one more section in this summary that caught my eye, and it is something we see increasingly in FDA documents: cybersecurity. The summary mentions Methinks employed a risk-based cybersecurity strategy.

Maya: Yes, they state they integrated cybersecurity into the design according to FDA guidance and Section 524B of the FD&C Act. This includes threat modeling, vulnerability assessment, third-party penetration testing, and a Software Bill of Materials, or SBOM.

Sam: Why is there so much focus on cybersecurity for an imaging tool?

Maya: Because this software sits directly on the hospital network. It ingests raw patient scans, processes them, and pushes notifications out to doctors. If this system gets compromised, it could either expose protected health information or be used as a backdoor for ransomware into the hospital's PACS system. The FDA wants to see secure update mechanisms and system logging in place.

Sam: That makes total sense. So, taking a step back, the conclusion of the summary states that Methinks CTA Stroke is substantially equivalent in technological characteristics, safety, and performance to the ContaCT device.

Maya: Exactly. They proved it can flag large vessel occlusions rapidly, with an overall sensitivity of 98.2% and specificity of 91.6% across 336 retrospective cases. And importantly, it adds the MCA-M2 vessel to the detection capabilities.

Sam: But the trade-off is knowing where the false positives live. If you have an older patient population with lots of chronic lesions or severe stenosis, you are going to see a dip in specificity, meaning more alerts that turn out not to be an acute occlusion.

Maya: Which brings us right back to the FDA's main rule for this class of device: it is a parallel workflow tool. It should not be used in-lieu of full patient evaluation or relied upon to make or confirm diagnosis.

Sam: That is a great place to wrap up our spotlight on this clearance. To recap, today we reviewed the FDA 510(k) summary for Methinks CTA Stroke, an AI triage tool designed to analyze brain scans for severe blockages and notify specialists in roughly 3.30 minutes.

Maya: We have linked the full public FDA document in the show notes so you can look at the subgroup performance tables yourself. And as always, a reminder that we are AI-generated voices, this podcast is for informational purposes only, and this is not medical advice.