CNNT-Net: a novel deep learning-based framework for estimating high-quality fiber orientation distributions from single-shell dMRI.

This study aims to introduce a novel deep learning-based network for the accurate reconstruction of high-quality fiber orientation distribution (FOD) images in the context of non-invasive imaging of white matter tractogram. The proposed network utilizes the angular correlation coefficient (ACC) as a threshold and employs a voxel selection module to categorize the voxels into two sizes. It integrates convolutional neural networks (CNN) and Transformers for estimating the FOD. Additionally, in instances where complex fiber configurations result in insufficient spherical harmonics (SH) coefficients estimated by the CNN below the threshold, the Transformer with two innovative attention mechanisms replaces the CNN's coefficients to enhance FOD estimation accuracy. Experiments conducted on the Human Connectome Project (HCP) dataset demonstrate promising qualitative and quantitative outcomes, showcasing the robustness of the proposed method in producing high-quality FOD images. The developed deep learning-based network presents a viable solution for enhancing the clinical applicability of high-quality FOD imaging, offering a reliable reference for non-invasive imaging of white matter tractogram.