Dictionary-Free MR Fingerprinting via Implicit Neural Representation

Magnetic Resonance Fingerprinting (MRF) enables rapid multi-parametric quantitative MRI but traditionally relies on discrete precomputed dictionaries, leading to quantization errors and limited flexibility. In this work, a Dictionary-Free Gradient-based Implicit Neural Representation (DFG-INR) framework is presented to learn direct mappings from spatial coordinates to tissue parameters (T1, T2, and PD) without any precomputed dictionary or labeled data. By integrating an INR with a fully differentiable Bloch simulator, DFG-INR enables end-to-end, physics-consistent optimization directly from acquired k-space data. The proposed method is validated using BrainWeb simulations, 3T high-field physical phantom experiments, and 5T ultra-high-field in vivo human brain data. Experimental results demonstrate that DFG-INR achieves superior quantitative accuracy, better anatomical detail preservation, and stronger noise robustness than state-of-the-art model-based MRF methods.