Protocol-Agnostic Q-Space Learning for Diffusion MRI Microstructural Parameter Reconstruction Using Diffusion Models

Advanced diffusion MRI (dMRI) models such as diffusion kurtosis imaging (DKI) and neurite orientation dispersion and density imaging (NODDI) provide rich microstructural information but require dense Q-space sampling and time-consuming computation. Learning-based approaches have been proposed for acceleration; however, their reliance on fixed acquisition protocols limits their generalization ability. We propose a diffusion-modelbased framework for protocol-agnostic reconstruction of multiple dMRI microstructural parameter maps from sparsely and heterogeneously sampled Q-space data. The method employs a denoising diffusion implicit model with an encoder-modulator-decoder architecture. A dual-branch encoder integrates anatomical priors from T1-weighted and b0 images with diffusion information aggregated by a Q-space resampler, mapping variable-length diffusion measurements into a unified latent representation. Diffusion-aware feature modulation and a multi-head decoder enable stable joint reconstruction of multiple parameters. Experiments on the Human Connectome Project dataset demonstrate that our method achieves PSNR above 24.86 dB and SSIM above 0.86 for all parameter maps across different numbers of diffusion-weighted images, consistently outperforming state-of-the-art learning-based approaches. Ablation studies further show that the proposed components contribute substantially, with the full model yielding the best reconstruction performance. These results indicate that our method provides an effective solution for protocol-agnostic dMRI parameter reconstruction in accelerated and heterogeneous acquisitions.