Controllable Cardiac Synthesis via Disentangled Anatomy Arithmetic

Acquiring annotated data at scale with rare diseases or conditions remains a challenge. It would be extremely useful to have method that controllably synthesizes images can correct such underrepresentation. Assuming proper latent representation, the idea of “latent vector arithmetic” could offer means achieving synthesis. A representation must encode fidelity input data, preserve invariance and equivariance, permit arithmetic operations. Motivated by ability disentangle into spatial anatomy (tensor) factors accompanying imaging (vector) representations, we propose framework termed “disentangled arithmetic”, in which generative model learns combine anatomical different when they are re-entangled desired modality (e.g. MRI), plausible new cardiac created target characteristics. To encourage realistic combination after step, localized noise injection network precedes generator. Our is used generate images, pathology labels, segmentation masks augment existing datasets subsequently improve post-hoc classification tasks. Code publicly available https://github.com/vios-s/DAA-GAN.