Computer Vision News - October 2022

39 Thanos and Hadrien and generate such counterfactuals and a practical application for diagnosing cardiac function with ultrasound.” Thanos and Hadrien are already considering the next steps and are keen to iron out a small flaw they have identified in their approach. “The causal framework is supposed to have variables which are completely independent of each other,” Hadrien points out. “In our case, the factual and something we call the confounder share some information. We’re thinking of ways to change this. It’s challenging and implies changing many things in the model, but that’s one of the flaws. It’s a work in progress for us.” Could we see the result at MICCAI in Vancouver next year? “We hope so!” Co-authors of this paper include experts in causality and cardiology who have guided Thanos and Hadrien on this topic. “When we were brainstorming how to approach the subject, the cardiologist told us that what we were proposing was interesting and novel,” Hadrien reveals. “They were keen to see what could be done and where it could be used. It was very encouraging.” Bernhard Kainz , Reader in the Department of Computing at Imperial and one of the paper’s co-authors, gives us his perspective: “I’m heavily biased, of course, but this work is novel because it’s the first MICCAI paper that shows how causality research and machine learning can be connected to provide new tools for diagnostic support. Doctors ask, ‘How would a patient’s scan look if clinical parameter X was different?’ We show several theoretical ways to learn This is a static image representing 2 key moments of the cardiac cycle, called the end-systole (ES) and end- diastole (ED) for the factual (=real world) echocardiogram on the left and the counterfactual (=fake, created by the model) echocardiogram on the right. The image shows that the model changed the real echocardiogram to lower the ejection fraction while retaining the anatomy of the patient in the original echocardiogram. ST OF ICCAI