Implicit Neural Representation-Based MRI Reconstruction Method With Sensitivity Map Constraints

Magnetic resonance imaging (MRI) is a widely used diagnostic tool, but its long acquisition time limits clinical efficiency. Implicit neural representation (INR) has recently shown great potential in scan-specific fast MRI reconstruction without requiring fully sampled training data. However, most existing INR-based methods overlook the structural properties of coil sensitivity maps and apply regularization only to the reconstructed image. To address this, we propose INR-CRISTAL, a joint image and sensitivity estimation framework that introduces explicit regularization on sensitivity maps to leverage their inherent smoothness. Additionally, we design a frequency-adaptive sinusoidal representation network (FA-SIREN) that enables dynamic frequency modulation at each layer, improving the adaptability to heterogeneous spectral components within complex signals. Crucially, our entire framework remains fully self-supervised. Experimental results demonstrate that INR-CRISTAL achieves more accurate sensitivity estimation and superior reconstruction quality, particularly under limited autocalibration signal (ACS) data and high acceleration rates.