MICCAI 2022 Daily – Tuesday

counterfactual from the data we observe? There are an enormous number of potential counterfactuals. In our case, because the echocardiograms and the videos were difficult to identify, we used a combination of theoretical constraints, plus we went to cardiologists and said, ‘Does this look reasonable to you?’ to inform our decision process. ” 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. This process is not unique to machine learning but used in many fields that feature causality and counterfactuals, such as econometrics and epidemiology. It is always important to be careful and strict about any assumptions, how they are made, and how they are implemented. Results could be wrong, but theoretical constraints and expert knowledge can help mitigate this. Did anything go wrong? Or did they see any surprising results? “ We were curious to see how the model would perform under extreme conditions, so we did produce some videos where, in real life, the patient would be dead! ” Thanos laughs nervously. “ The video itself was believable 11 DAILY MICCAI Tuesday Thanos and Hadrien