MICCAI 2022 Daily – Tuesday

they needed ground truth to compute an error on the output of both branches. For the factual branch, that was not difficult because they had a video input and wanted the network to reconstruct that input. However, for the counterfactual branch, they had no way of generating what did not exist. This shows 3 triplets of echocardiograms: the input (confounder) and both outputs (factual and counterfactual). It demonstrates that the anatomy is preserved while the ejection faction changes (the amplitude of the movement is different). “ That was the hardest part – we had to find a trick, ” he says. “ The solution we found was not to use videos as our ground truth, but instead use two neural networks to compute the loss on metrics that we wanted the network to learn. We wanted the counterfactual branch to generate a video that looked real and a video that had a different ejection fraction. We could train a network to ensure that the video produced looked like a real echocardiogram, and we knew how to make a network that would evaluate the ejection fraction in the counterfactual video, so we could backpropagate the loss through this expert network to enforce the ejection fraction. ” On the theoretical side, the problem was that, by definition, the counterfactual was an alternative world they did not have access to but wanted to contemplate. Even figuring out a video would look a certain way, they had no way to tell if it was true. They had to apply some stringent conditions to allow them to make such claims. “ Those conditions are essentially called identifiability conditions , ” Thanos explains. “ The question here is, can we identify the correct counterfactual 10 DAILY MICCAI Tuesday Poster Presentation