Spatial-Sequence Joint Modeling for Accelerated Multi-Model Diffusion MRI Microstructure Estimation

This study aims to develop an efficient and high-fidelity deep learning framework for accelerated multi-model diffusion MRI microstructure estimation using sparsely sampled q-space data.

We propose a shared-encoder, distinct-decoder framework. A dual-branch architecture within the shared encoder integrates convolutional neural network (CNN)-based spatial modeling for local tissue structures and Mamba-based sequence modeling for global structural priors. The fused features form a shared latent representation that regularizes the highly subsampled estimation process. Distinct decoders extract model-specific features from this latent space, enabling diverse parameter estimation across biophysical models. A joint estimation loss with tunable weights balances the multi-model learning objectives.

Extensive experiments on the Human Connectome Project (HCP), Alzheimer's Disease Neuroimaging Initiative (ADNI), and Tiantan clinical dataset demonstrate consistent improvements over six state-of-the-art methods under multiple q-space subsampling settings. The proposed framework supports 4.5×-27× q-space acceleration while achieving 3×-9× faster inference. Additional experiments further support the robustness and cross-dataset adaptability of the proposed framework.

The proposed framework enables accurate and reliable joint estimation of multi-model dMRI microstructural parameters from accelerated acquisitions and provides a practical solution for fast microstructural imaging.