Denoising Knowledge Transfer Model for Zero-Shot MRI Reconstruction

Though fully-supervised deep learning methods have made remarkable achievements in accelerated magnetic resonance imaging (MRI) reconstruction, the fully-sampled or high-quality data is unavailable in many scenarios. Zero-shot learning enables training on under-sampled data. However, the limited information in under-sampled data inhibits the neural network from realizing its full potential. This paper proposes a novel learning framework to enhance the diversity of the learned prior in zero-shot learning and improve the reconstruction quality. It consists of three stages: multi-weighted zero-shot ensemble learning, denoising knowledge transfer, and model-guided reconstruction. In the first stage, the ensemble models are trained using a multi-weighted loss function in k-space, yielding results with higher quality and diversity. In the second stage, we propose to use the deep denoiser to distill the knowledge in the ensemble models. Additionally, the denoiser is initialized using weights pre-trained on nature images, combining external knowledge with the information from under-sampled data. In the third stage, the denoiser is plugged into the iteration algorithm to produce the final reconstructed image. Extensive experiments demonstrate that our proposed framework surpasses existing zero-shot methods and can flexibly adapt to different datasets. In multi-coil reconstruction, our proposed zero-shot learning framework outperforms the state-of-the-art denoising-based methods.