Unsupervised MRI Reconstruction via Zero-Shot Learned Adversarial Transformers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Supervised MRI reconstruction requires matched undersampled and fully-sampled data for training.
• Deep image prior methods reduce supervision but use suboptimal convolutional architectures.
• Untrained priors may lead to suboptimal performance in MRI reconstruction.

Purpose of the Study:

• To introduce a novel unsupervised MRI reconstruction method, SLATER (zero-Shot Learned Adversarial TransformERs).
• To address limitations of existing methods in capturing long-range relationships and improving MRI reconstruction performance.
• To develop a method that reduces supervision requirements while maintaining high-quality reconstruction.

Main Methods:

• SLATER utilizes a deep adversarial network with cross-attention transformers for MRI reconstruction.
• An unconditional network is pre-trained for unsupervised generative modeling to learn an MRI prior.
• Zero-shot reconstruction is performed by incorporating the imaging operator and optimizing the prior for data consistency.

Main Results:

• SLATER demonstrates superior performance in unsupervised MRI reconstruction.
• The method effectively learns a high-quality MRI prior through unsupervised generative modeling.
• Experiments on brain MRI datasets show significant improvements over state-of-the-art unsupervised methods.

Conclusions:

• SLATER offers a powerful unsupervised approach for MRI reconstruction.
• The use of adversarial transformers enhances the capture of long-range dependencies in MRI data.
• This method advances the field of medical image reconstruction by reducing supervision needs and improving accuracy.