Method and apparatus for using generative adversarial networks in magnetic resonance image reconstruction

A method of training a generator network, comprising:

• determining, over each iteration of a plurality of iterations, an adaptive combined loss comprising a pixel-wise loss and an adversarial loss, wherein the pixel-wise loss corresponds to a direct comparison between a generated image generated by the generator network and a ground truth image, and the adversarial loss corresponds to an estimated comparison between the generated image and the ground truth image estimated by a discriminator network, wherein the adversarial loss is adaptively adjusted based on a ratio between the pixel-wise loss and the adversarial loss such that the ratio is below a threshold;
• and
• updating, during each iteration, the generator network based at least in part on the adaptive combined loss such that, in subsequent iterations, the updated generator network is more likely to generate a subsequent generated image that better represents a subsequent ground truth image than the generated image represented the ground truth image.