Brain age from white matter fiber orientation: 3.54-year mean error across 692 people, and predictions rotate with the input

The input to brain age prediction shifts from diffusion tensor scalar metrics to spherical harmonic coefficients of white matter fiber orientation distributions, so orientation information reaches the model. Many existing prediction models use diffusion tensor imaging scalar metrics, which omit orientation information. Evaluated on 692 participants aged 36-90 from the Human Connectome Project Aging dataset, the combined framework outperformed the evaluated DTI-based models and conventional CNNs.

The network operates directly on spherical harmonic coefficients and preserves their transformation properties under rotation. The method is an SE(3) equivariant CNN, a convolutional network equivariant to 3D rotations; conventional CNNs do not explicitly respect the geometric structure of orientation-dependent features. Averaging predictions from five cross-validation models yielded a mean absolute error of 3.54 years and a Pearson correlation of 0.958.

Model sensitivity can be localized tract by tract and moves with age band. Masking each segmented tract in turn placed sensitivity peaks in different pathways at different ages. Peaks were in frontal projection pathways at 36-44 years, the fornix at 45-54, the callosal isthmus at 55-64, the middle cerebellar peduncle at 65-74, and the callosal genu at 75-90.

The evaluation covers 692 participants aged 36-90 from the Human Connectome Project Aging dataset, compared against the evaluated DTI-based models and conventional CNNs.

The performance advantage depended on combining WM-FODs with equivariant processing; prediction accuracy was highest at maximum spherical harmonic order Lmax = 2 and declined at higher orders.