Learning-based MRI Reconstruction Method with Coil Sensitivity Estimation and Prior Adaptation

Magnetic Resonance Imaging (MRI) reconstruction from undersampled multi-coil k-space data is a challenging inverse problem, typically relying on fixed priors such as precom puted coil sensitivity maps. To improve adaptability, we propose a novel framework named Learning-based Coil sensitivity Estimation and Prior Adaptation (LCEPA). LCEPA formu lates reconstruction as a bilevel optimization, using a Joint Deep Equilibrium (Joint DEQ) model in the inner loop to simultaneously estimate images and coil sensitivity maps, while the outer loop adaptively fine-tunes image and coil priors using supervised data. Experiments demonstrate that LCEPA surpasses state-of-the-art methods in terms of PSNR and SSIM, showcasing robust generalization to varying acceleration rates and imaging modalities. Our approach thus provides an effective and flexible solution for accelerated MRI reconstruction in clinical scenarios.