Current approaches for synthetic multi-contrast MRI involve deep networks trained to synthesize target-contrast images from source-contrast images in fully-supervised protocols. Yet, their performance is undesirably circumscribed to training sets of costly fully-sampled source-target images. For practically advanced multi-contrast MRI synthesis accelerated across the k-space and contrast sets, we propose a semi-supervised generative model that can be trained to synthesize fully-sampled images using only undersampled ground-truths by introducing a selective loss function expressed only on the acquired k-space coefficients randomized across training subjects. Demonstrations on multi-contrast brain images indicate that the proposed model maintains equivalent performance to the gold-standard fully-supervised model.
This abstract and the presentation materials are available to members only; a login is required.