MICCAI 2021 Daily – Wednesday

“ We used a probabilistic programming language called Pyro to implement a structural causal model, which is essentially a Bayesian network with causal interpretations, ” Jacob explains. “ The model is composed of a bunch of variables with cause-effect relations. I’ve used expert knowledge to define those relations between the variables, such as age, biological sex, brain volume, and lesion burden, and the image itself. We learn the cause-effect relationship between variables by backpropagation using the evidence lower bound as a loss function. This trains all components of the model at the same time. The two major components of our model are normalizing flows and variational autoencoders. The normalizing flows are used for covariates, and the variational autoencoder is used for the high-dimensional medical image. ” The work builds on previous works by Pawlowski, Castro, and Glocker , who themselves built on the work of Judea Pearl , a founding figure in computer science for causal inference. It is extending a causal model to a specific disease and demonstrates that it is possible to intervene on certain variables and generate counterfactuals which appear to be what you would expect. 13 DAILY MICCAI Wednesday Jacob Reinhold This picture shows four images generated by ancestrally sampling the trained structural causal model. The picture shows the causal relationship between covariates in the model.