ESOC 2026 | Improving emboli detection and reperfusion assessment with a novel image-processing approach

As part of my PhD in a broader sense, we look at image processing for the treatment of stroke and all the imaging around that, so the CT and mainly the digital subtraction angiography is what I look at with all these different imaging modalities. And in digital subtraction angiography, this is an imaging that’s made in a very acute setting because the patient is being treated. So you have to go fast to the intervention room and then you’re trying to remove the clot with the catheter. And so this means there’s a lot of room for human error in the interpretation of these images. So one thing that you do is you remove the clot and then you have, so you first have an image before a clot removal where you make a digital subtraction angiography. You inject contrast and you can see the blood vessels. You have one before you pull out the clot. So we call that the pre-EVT DSA, so before endovascular thrombectomy imaging. And then you also have one after and you can compare those. And this comparison is done usually by the radiologist who performs the procedure and they look how successful they are. They look how much clot has been removed and they give a score to this called the TICI score. And we realized that this process for us from the image processing side that we can automate this somewhat, so we can actually try to align these blood vessel trees so the image before and image after and then you can look at the differences and this allows you to then do image processing and kind of quantify the difference. So that’s what we’ve done. We’ve aligned it. We’ve aligned 40 pairs of these pre- and post-images and we asked radiologists to do a rating of the TICI score, so to rate the reperfusion score. And we’ve done this without first to do a control. We do it without just giving them the normal imaging as they would do it normally, and then we did the same but we did it with our overlay system that we made, so we lay the images on top of each other and we can highlight what the differences are in the vessels. And this was actually interesting to see because we saw that there is an increase in the agreement between the radiologists for the TICI scoring. Because this TICI scoring, there’s quite some differences in how, like I said, there’s human error involved. Different humans will look at the same images and give it different scoring. But if we give them additionally this overlay, the agreement increases somewhat.

And also interesting as a kind of a secondary finding, we see that this image can also highlight, this overlay can highlight when there’s a different vessel that is now disappearing. So we are focused, of course, you’re trying to remove the clot, so you’re opening a blood vessel, and there are some vessels appearing in that area, hopefully, if you’re successful. But during this procedure, you are with a wire in the blood vessels, you pull some clot back, a piece can break off and get into a different vessel. And this can be picked up by this tool, because you can see, like, you overlay the vessels, and you see that hey, there was a vessel here before and it’s now missing. So this is another I think relevant potential avenue to go. So this is of course a relatively small study. We had only 40 cases, but it seems that the radiologist can find more often these new infarctions with this overlay tool. So that’s also something we think can be clinically relevant, and we can pursue this further if we get more data and evaluate it in a larger setting, because the radiologist can be very focused on their own area that they just opened the blood vessel and then you can miss that there’s now another vessel somewhere else that has disappeared. And this is especially, we think, also relevant because stroke patients, they often come at night, under different settings, when you are, when the interventionalist is not, let’s say, maybe cognitively optimally performing in that moment. So then it’s very nice to be able to have this kind of tool, this kind of helper tool that can do this overlay and show the differences in vessels and based on that give some additional support to the interventionist.

Yeah, so that’s hopefully what we can move into the direction of this tool. But of course, it’s not without limitations. And I think the main issue here is this DSA. You’re making a two-dimensional image. It’s like an X-ray, real-time X-ray, where you inject the contrast. But if the patient’s head moves, like they tilt their head, then you get a different two-dimensional image when you look again. And if there’s too much difference between these two images, then our relatively simple image processing method cannot align the images anymore. So that’s still a big challenge that we hope to investigate more and overcome. And we are looking now into, so this is kind of a pilot where we did it in a small setting. We look now to develop an artificial intelligence model that can be trained to really be robust against these perturbations, against these movements of the head. So that will be the next step, and then hopefully we can evaluate it on a bigger set.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.