Frequency-enhanced fiber orientation distribution super-resolution network with application on Alzheimer's disease.

BACKGROUND Diffusion MRI enables noninvasive assessment of white matter microstructure, but the accuracy of fiber orientation distribution (FOD) reconstruction remains limited by inter-voxel incoherence and insufficient modeling of high-order spherical harmonic dependencies. Existing FOD angular super-resolution approaches often lose orientation information in complex fiber configurations, and their applicability in disease-specific scenarios has been underexplored.

NEW METHOD We introduce the Cross-Domain Attention FOD Network (CDAF-Net), which integrates a Frequency-enhanced Aggregated Spatial-Channel Attention module to jointly model frequency-domain, spatial, and channel features. The frequency branch captures long-range dependencies, while spatial and channel attentions highlight local critical structures, enabling efficient and accurate angular super-resolution. The apparent fiber density (AFD) derived from CDAF-Net was further evaluated using one-way ANOVA across Alzheimer's disease, mild cognitive impairment, and cognitively normal groups.

RESULTS CDAF-Net yielded the most accurate reconstructions, achieving the closest alignment with reference FODs and showing clear advantages in regions with complex fiber crossings. Clinically, CDAF-Net-derived AFD exhibited bilateral fornix reductions consistent with disease progression and demonstrated significant group differences, with larger effect sizes and higher sensitivity than baseline approaches, while other white matter tracts did not reach significance.

COMPARISON WITH EXISTING METHODS Compared with state-of-the-art FOD angular super-resolution models, CDAF-Net produced lower orientation errors, better preserved directional information in multi-fiber regions, and provided more sensitive detection of microstructural alterations related to early Alzheimer's disease.

CONCLUSIONS CDAF-Net improves FOD reconstruction fidelity and enhances sensitivity to early neurodegenerative changes, supporting its potential as a cost-effective alternative to multi-shell HARDI for early screening and clinical research applications.