AutoSyncoder: An Adversarial AutoEncoder Framework for Multimodal MRI Synthesis

The ability to generate multiple contrasts for the same patient is unique about MRI and of very high clinical value. In this work, we take up the problem of modality synthesis in multimodal MRI and propose an efficient, multiresolution encoder-decoder network trained like an autoencoder that can predict missed inputs at the output. This can help in avoiding the acquisition of redundant information, thereby saving time. We formulate and demonstrate our proposed AutoSyncoder network in a GAN and cyclic GAN setting, and evaluate on the BRATS-15 multimodal glioma dataset. A PSNR ranging between 29 to 30.5 dB, and SSIM over 0.88 is achieved for all the modalities, with simplistic training, thereby establishing the potential of our approach.