A Novel Deep Learning Approach for Tissue Microstructure Estimation Using Spherical Mean Technique

Multi-shell diffusion magnetic resonance imaging (dMRI) models connect tissue organization with the observed dMRI signals to infer tissue microstructure. Given the highly non-linear nature of these models, several deep network based approaches have been proposed to improve the efficiency and accuracy of tissue microstructure estimation. However, these approaches rely on dMRI data with dense q-space sampling, which can lead to high computational costs. In this work, we propose a novel deep learning framework, termed DeepSMT, which utilizes the spherical mean technique (SMT) to directly estimate NODDI and DKI microstructure. The network comprises an inter-voxel block and a multi-shell feature block that can capture the feature relationship of the average signal between inter-voxels and integrate features from each shell. We validate the effectiveness of our method using the Human Connectome Project dataset. Overall, the DeepSMT has better performance compared to other methods in estimating tissue microstructure.