Prof. Stefan Niehues, MD, MHBA, Prof. Peter Chang, MD, Prof. Kenneth V. Snyder MD, PhD, FAANS, Benoit Doche de Laquintane, MD

"The algorithms aren’t perfect and there's much room for improvement. They work well when the data is similar to the types of images used for training. However there are issues where imaging protocols aren’t standard. So making models more generalizable for images from any manufacturer is a key issue.
Similarly, machines evolve and scanners change. So as the technology for the scanners improves, our AI doesn’t currently adapt to that technology. Developing ways for the AI system to improve itself on the fly, understand new imaging protocols, and get better is another very active area of research we're working on," said Peter Chang.
"So a simple software update on your machine may affect algorithm precision?" commented Stefan Niehues. "That's correct. An update that seems trivial to us ends up affecting performance. We’re researching ways to detect that without needing a human.
It’s part of the secret to taking AI to the next level.," responded Peter Chang.

"We're currently engaged in those trials. I think we underestimate the power and importance of AI clot-busting drugs into that distal territory and tissue. There are also novel sub-one-and-a-half millimeter devices out there. However there are challenges when you get very distal. First of all, we underestimate the importance and devastation of a small vessel stroke in the wrong location.
If you have receptive aphasia that completes itself you are lost to the planet. You are unrehabable and so it’s important to be very aggressive about certain distal strokes just as it is around an LVO.
And I think what's also important is crossing the clot before you attempt to remove it and looking at the distal vasculature. As once you get far out like that you can end up with a whole distal territory that's filled with clot. Just trying to pull a plug out does you no good because there's a clot all the way out.
Intra-arterial TPA is another area of research that we're looking at right now and exploring," added Kenneth V. Snyder. "That's right and from an AI perspective, we're starting with M1 and proximal M2 because it's easy. It’s something that generalizes well to many different hospitals. However, as soon as folks like Professor Snyder tell us it's important, it's our job to make those algorithms. There's no reason the Deep Learning systems can't do this. We just need to work on it," commented Peter Chang. "What will probably start to revolutionize the field is if we get away from non-diffusible tracers and bring back xenon and other things like gas diffusable tracers. As now we'll have an exact metabolic measure that will be much more reliable for our patient selections. Again, that type of advancement would revolutionize stroke," commented Kenneth V. Snyder. "Theoretical limitations are really not the problem. As long as you have relevant data to train the model you'll be fine. Our limitation in making the first generation of algorithms is the fact that M2 and M3 occlusions just aren't that common, or at least they're not commonly reported in the radiology workflow. So we don't have a lot of those examples to feed our algorithms.
As you get to the smaller vessels, as you might imagine, the appearance also changes dramatically – more so based on manufacturing protocol, venous contamination, and whether you have a good arterial bolus…. right?
These things add some complexity and variability to the process. However by and large, if you have enough training data, an algorithm can learn the relevant patterns. So it's not a problem with the technology. We just we have to get the data and build that algorithm," added Peter Chang. "Patients with acute stroke rarely lie straight and stay still, which is more horrible for MRI imaging than CT. So it would be great to get algorithms that work on moving patients," concluded Stefan Niehues.

"I think you will do a CT in this case because time is brain and you need the quickest solution. I don't think you will push an MRI," said Doche de Laquintane. "I agree completely added Kenneth V. Snyder. "In the rare situation where you can get an MRI very quickly, our standard, regardless of stroke or another disease process, would be something like an X-ray just to exclude metal implants and other things like that," added Peter Chang. "Yeah, the last thing you want to do is harm the patient," commented Stefan Niehues.

"I don't have much experience in the emergency unit, but I imagine you would go for the quickest solution. I think most people would use a CT with perfusion. MRI is good but it takes around seven minutes. Again, time is brain," said Doche de Laquintane. "I'd like to add that there's an inherent difficulty in that perfusion imaging in MR is non-linear in its analysis and CT is linear in terms of its signal. That adds another dimension of complexity to the thresholding and analytics of the assessment. Although MR is important and helpful, especially for small strokes or diagnosis, there is a clear role right now for CT perfusion.
The other point I want to bring up is the importance of not intubating a patient if you don't need to. These patients are auto-regulating when they come in. They're vasodilating and driving their own blood pressure. Their brain is deciding how much blood it needs. The minute you intubate a patient, you're now dependent on an anesthesiologist for blood supply and we've learned the hard way that is not ideal," added Kenneth V. Snyder. "I essentially agree. I know of maybe two or three centers in the whole of the US that start with an MR based triage protocol and it's only because they have MRI-dedicated stroke imaging in the ER. So they can roll the patient straight in and get that done right away.
Even in centers that do have that capacity, there are the complications just mentioned. So it’s a CT first, by and large," concluded Peter Chang.

"Yes, it potentially can. Our algorithm reorients the patient as soon as the images are acquired. We take the raw 3D volume and make sure everything is anatomic before analyzing the studies for that reason. I don't know what other AI solutions do, but there will be problems if they don’t compensate," stated Peter Chang.

"It depends on what you're looking for. You’ll miss subtle hemorrhage or very subtle findings on CT if you open up the field too significantly. It's not just the matrix size that's changing. Your dose calculations completely change and you need to penetrate more tissue. So it complicates the situation. That being said, some things don't require a lot of detail. Our perfusion data, for example, is acquired at a very coarse resolution. It's interesting. I'd be curious to know what the exact amount of degradation would be," said Peter Chang.
"Building off that comment, I think it’s important for people to understand that a lot of perfusion imaging is for people with severe NIHs that you're going to potentially take for treatment and that are likely to have an LVO. For your low NIH presentations, those will all go to MR and you'll be able to get that and do those comparisons," commented Kenneth V. Snyder. "Thank you so much for the impressive cases and discussions. I would also like to thank all participants for attending and thank Canon Medical for making this webinar possible. Please keep an eye on the Canon Medical website for details on upcoming talks," added Stefan Niehues. //

See more topics in this webinar.