Deep Learning Empowered Microstructure Codebook: New Paradigm for Multi-Parameter Tissue Characterization Estimation

Diffusion MRI (dMRI) enables the examination of microstructural profiles and tissue changes using specific microstructural modeling, but it requires long acquisition times and dense q-space sampling. Current deep learning-based methods are also limited by their inability to generalize across protocols and extend to new microstructural indices. This work introduces a novel framework that addresses these limitations by learning a microstructural codebook, facilitating accurate, rapid, and multi-parameter microstructure imaging. Our approach integrates the spherical mean technique (SMT) with a hybrid Mamba-CNN architecture and learnable tissue-compartment kernels, effectively capturing multiscale spatial dependencies while linking spherical mean signals to biophysical microstructure models. This design enhances both interpretability and adaptability, enabling robust estimation of 24 microstructural metrics derived from 8 widely used biophysical diffusion models, even under undersampled acquisition conditions. Notably, the framework demonstrates strong generalization across diverse acquisition protocols and enables seamless adaptation to novel microstructural indices with minimal fine-tuning, underscoring its flexibility and practical utility. Extensive experiments on multiple datasets confirm the method's superior accuracy, generalization, and transferability. This work presents a codebook-driven framework for microstructure imaging that bridges biophysical modeling and deep learning to enable more interpretable and adaptable dMRI analysis. The code is available at https://github.com/1nlandempire/Microstructure-codebook-imaging.