MICCAI 2018 Daily - Wednesday

ThinkSono 13 Wednesday Antonios talks about some of the challenges they have faced so far: “First of all, the model has to work real time – which is in contrast with most of the models that we see here at MICCAI – because the clinician has to see in real time the result of the diagnosis. That was one challenge that we had to make the model very lightweight but also have a very high accuracy. The second challenge, and a very big one, is the lack of data. Specifically, for DVT, there was a lack of data and a lack of correctly acquisitioned data. Also, for labelled data. That’s another completely different thing. We have to develop a whole protocol on how to label this data to be able to solve the problem in a way that a radiologist would.” Sven explains how they expect their solution to find a place in the real world: “You would have to look at how DVT is diagnosed today. That’s a very long pathway. A patient maybe has symptoms, goes to their general practitioner or emergency department, and just by catching these symptoms, doctors are aware that this can be an emergency and it has to be treated immediately. They are sent off to the radiology department to then do a proper scan. Now, radiologists have very scarce resources and they need high-end equipment, and it’s always busy there, so they try to screen patients beforehand to not go into the scan. To give you an example, in London, the Guy’s and St Thomas’ Hospital sees 1,400 patients a year, but only 70 of those are positive. Similar hospitals we screen are in the same range, so there needs to be an improvement in the process to screen out patients who don’t have DVT. We need to place that in the emergency department to be used by emergency doctors and nurses.” In terms of next steps, they are looking to ensure the model has a consistent accuracy across all different vendors of ultrasound probes. Also, to make it very lightweight so that it can even run in very small, portable devices. Finally, Sven says there are different ways to present the output of these models to the user. It’s a part that can’t be underestimated, because if the user doesn’t know how to use the tech, the results can vary heavily. He says they have to really investigate and be crystal clear on how to use it in practice. He says their attitude towards this work has been to be very problem- centric and to aim to really make a difference for the clinicians, and an impact in the clinic. They aim to see their application in a hospital by 2020. Antonios (left) and Sven