TOPLINE:

A novel hierarchical deep learning framework successfully performed direct Desikan-Killiany (DK) brain parcellation using only diffusion MRI (dMRI)-derived data, achieving high structural segmentation accuracy and robustness without relying on anatomic scans or subject-specific cross-modality registration, a study showed.

METHODOLOGY:

• Researchers developed a sequential two-stage hierarchical deep learning framework for DK parcellation of the brain using only dMRI-derived input parameters including trace (T), fractional anisotropy (F), sphericity (S), and maximum eigenvalue (E1).
• The framework was trained, validated, and tested in a 50:30:20 ratio on a total of 100 young, healthy adults from the Human Connectome Project (HCP) dataset (mean age, 29.1 years; 46 men).
• The first stage performed coarse parcellation into seven broad anatomic regions, and the second stage refined segmentation into 101 detailed DK atlas labels using five dedicated three-dimensional convolutional neural networks.
• External validation was conducted on 214 participants from the Consortium for Neuropsychiatric Phenomics (CNP) dataset (mean age, 33.2 years; 114 men) including healthy control individuals and patients with schizophrenia, bipolar disorder, and attention-deficit/hyperactivity disorder.
• Primary outcome measures included Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance on HCP data, with relative SD (RSD) computed for diffusion-derived measures within each parcellated region.

TAKEAWAY:

• The hierarchical framework achieved a mean DSC of 82.09% (SD, 0.05) on the HCP dataset, with statistical significance after Bonferroni correction (P = .00468).
• The optimal four-parameter input combination of T + F + S + E1 yielded a mean DSC of 76.52% in two-dimensional screening, outperforming previously used parameter configurations, with statistical significance after correction.
• On the CNP dataset, the proposed models achieved lower RSD values than anatomic MRI registration across all three diffusion measures, showing statistically significant method effects for fractional anisotropy (P < .0001), mean diffusivity (P < .0001), and sphericity (P < .0001).
• The new model reduced class imbalance using the maximum-to-median voxel-count ratio, which decreased from 43.6 in the original 101-label problem to 1.2 in the seven-category coarse stage; improvements were observed in both large structures and smaller regions such as the caudate.

IN PRACTICE:

"On the CNP dataset, where reliable voxel-wise diffusion-space DK reference labels are not available, the lower RSD values provide label-free evidence that the proposed framework produces more homogeneous parcellations under heterogeneous acquisition conditions, while remaining complementary to the supervised evaluation on HCP," the authors wrote.

"Overall, the proposed method represents a step toward practical dMRI-based brain parcellation by avoiding the need for anatomical MRI and subject-specific anatomical-to-diffusion registration at inference time," they added.

SOURCE:

This study was led by Yousef Sadegheih, University of Regensburg, Regensburg, Germany. It was published online on June 03, 2026, in Scientific Reports.

LIMITATIONS:

The study was limited by model training on the HCP dataset consisting primarily of young, healthy adults; the lack of reliable voxel-wise DK reference labels in the diffusion space for the external validation dataset such as CNP; the inability to apply susceptibility-induced echo-planar imaging distortion correction due to unavailable auxiliary acquisitions; and the restriction of input parameters to a single b = 1000 s/mm² diffusion tensor shell.

DISCLOSURES:

This study received support from Erlangen National High-Performance Computing Center funding provided by federal and Bavarian state authorities, as well as from the German Research Foundation. The authors reported having no relevant conflicts of interest.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.