Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3D deep generative models

We describe Countersynth, a conditional generative model of diffeomorphic\ndeformations that induce label-driven, biologically plausible changes in\nvolumetric brain images. The model is intended to synthesise counterfactual\ntraining data augmentations for downstream discriminative modelling tasks where\nfidelity is limited by data imbalance, distributional instability, confounding,\nor underspecification, and exhibits inequitable performance across distinct\nsubpopulations. Focusing on demographic attributes, we evaluate the quality of\nsynthesized counterfactuals with voxel-based morphometry, classification and\nregression of the conditioning attributes, and the Fr\\'{e}chet inception\ndistance. Examining downstream discriminative performance in the context of\nengineered demographic imbalance and confounding, we use UK Biobank magnetic\nresonance imaging data to benchmark CounterSynth augmentation against current\nsolutions to these problems. We achieve state-of-the-art improvements, both in\noverall fidelity and equity. The source code for CounterSynth is available\nonline.\n