IMJ-PLUS: Implicit Representation for Dynamic MRI and Coil Sensitivity Joint Reconstruction Using Low-Rank PLUS Sparse Regularization

Dynamic magnetic resonance imaging (MRI) must balance spatial and temporal resolution, a challenge that becomes more pronounced in free-breathing scenarios due to unpredictable motion. Existing deep learning (DL) methods rely on fully sampled data and assume static coil sensitivity maps (CSMs), which can lead to poor generalization and error propagation, particularly in dynamic, free-breathing imaging. To address these limitations, we introduce IMJ-PLUS, an unsupervised reconstruction framework based on Implicit Neural Representation (INR). IMJ-PLUS decomposes the image sequence into low-rank and sparse components, capturing both the static background and dynamic regions. Additionally, it jointly estimates time-varying CSMs by modeling them as continuous spatiotemporal functions. Unlike supervised methods, IMJ-PLUS optimizes these functions directly from undersampled k-space data acquired from a single subject, without relying on external training datasets. Our results show that IMJ-PLUS consistently outperforms compressed sensing, DL, and INR-based approaches across different acceleration factors (AFs). At extremely high AFs (up to 82×), it delivers PSNR gains of 1.02-8.32 dB. IMJ-PLUS also demonstrates significant performance gains in free-breathing scenarios.