2021/05/14 by Yılmaz Korkmaz, Korkmaz, Yilmaz, Salman UH Dar +7 · 4 citations
Medicine · #Advanced MRI Techniques and Applications #Medical Imaging Techniques and Applications #Radiomics and Machine Learning in Medical Imaging
paper · pdf · doi:10.48550/arxiv.2105.08059
Supervised reconstruction models are characteristically trained on matched\npairs of undersampled and fully-sampled data to capture an MRI prior, along\nwith supervision regarding the imaging operator to enforce data consistency. To\nreduce supervision requirements, the recent deep image prior framework instead\nconjoins untrained MRI priors with the imaging operator during inference. Yet,\ncanonical convolutional architectures are suboptimal in capturing long-range\nrelationships, and priors based on randomly initialized networks may yield\nsuboptimal performance. To address these limitations, here we introduce a novel\nunsupervised MRI reconstruction method based on zero-Shot Learned Adversarial\nTransformERs (SLATER). SLATER embodies a deep adversarial network with\ncross-attention transformers to map noise and latent variables onto\ncoil-combined MR images. During pre-training, this unconditional network learns\na high-quality MRI prior in an unsupervised generative modeling task. During\ninference, a zero-shot reconstruction is then performed by incorporating the\nimaging operator and optimizing the prior to maximize consistency to\nundersampled data. Comprehensive experiments on brain MRI datasets clearly\ndemonstrate the superior performance of SLATER against state-of-the-art\nunsupervised methods.\n