Free-breathing dynamic MRI reconstruction via joint time-dependent coil sensitivity estimation using implicit neural representation.

Dynamic magnetic resonance imaging (MRI) often encounters a trade-off between spatial and temporal resolutions due to slow data acquisition, particularly in free-breathing MRI scans where unpredictable motion complicates imaging. Accelerated dynamic MRI techniques based on highly undersampled k-space data have emerged as a promising solution to enhance spatiotemporal resolution. However, current supervised deep-learning reconstruction methods rely on fully-sampled training data and often struggle to generalize across diverse acquisition conditions. Moreover, these methods typically assume static coil sensitivities, overlooking motion-induced variations that can introduce errors in free-breathing dynamic MRI reconstructions. To overcome these limitations, we propose IMJ-PLUS, an unsupervised learning method for dynamic MRI reconstruction that leverages implicit neural representation (INR). Specifically, IMJ-PLUS models the image sequence as two continuous functions of spatiotemporal coordinates, decomposing it into background and dynamic components that are explicitly regularized using low-rankness plus sparsity (L+S) constraints. Additionally, IMJ-PLUS dynamically estimates time-varying coil sensitivity maps by representing them as continuous functions, enabling dynamic adaptation to motion. By leveraging the expressive power of INR, the effectiveness of L+S regularization, and joint dynamic coil sensitivities estimation, IMJ-PLUS outperforms the compared state-of-the-art methods in an unsupervised manner. Remarkably, under extreme undersampling conditions (acceleration factor up to 49.2  × ), IMJ-PLUS achieves substantial peak signal-to-noise ratio improvements, ranging from 1.2 to 7.7 dB. These results highlight IMJ-PLUS's potential to significantly accelerate dynamic MRI acquisitions while maintaining high image quality, even in the challenging context of free-breathing imaging. The source code is available at: https://github.com/AMRI-Lab/IMJ-PLUS.