KOMET: Kspace Optimal Masking for Efficient Training of Zero-shot MRI Reconstruction

Motivation: Zero-shot learning shows promise in various domains but remains underexplored in MRI reconstruction due to inherent challenges in recovering images without supervision of fully sampled data. Goal(s): To develop a stable zero-shot MRI reconstruction framework eliminating the need for fully sampled reference data. Approach: We propose a dynamic k-space masking strategy inspired by masked image modeling, coupled with joint optimization of image reconstruction and coil sensitivity estimation. Results: KOMET shows robust performance across various undersampling patterns (4×, 8× reduction with Gaussian and uniform masks) on FastMRI knee dataset, outperforming traditional parallel imaging methods and recent zero-shot approach with 2dB PSNR improvement. Impact: KOMET establishes a novel framework for stable zero-shot MRI reconstruction by adapting masked modeling to k-space domain. This advancement enables robust acceleration without fully sampled reference data, paving the way for broader clinical application of accelerated MRI.