Automatic sub-bundle tractometry reveals localized microstructural variations in complex white matter topographies

1 Introduction

• A single conventional bundle profile aggregates measurements along a representative streamline and assumes uniform microstructure across the bundle cross-section.In anatomically complex bundles, localized group differences may be diluted or missed.Source: §1 · pp. 2–3
• Earlier within-bundle clustering required bundle-specific distance-threshold tuning, and its benefit over full-bundle profiling had not been assessed.Whole-brain atlas clustering offers another route, but its effectiveness depends on atlas quality.Source: §1 · p. 3
• Automatic clustering within an existing anatomical bundle atlas is designed to distinguish geometric variability, including crossing and fanning patterns.The framework is evaluated using FA and IFW, with tests of microstructural heterogeneity, age sensitivity, and clinical group differences.Source: §1 · pp. 3–4
• Depression is a useful test case because reported white-matter effects are subtle and their within-bundle spatial distribution remains poorly characterized.Prior reports describe modest effects and note that observed effects may be driven by medication status and comorbidities.Source: §1 · pp. 3–4
• Late-life depression is especially relevant because aging affects white matter non-uniformly, including frontal and superficial regions where bundles are highly fanned.This motivates testing the framework in a setting where age-related changes intersect with complex tract geometry.Source: §1 · p. 4
• The evaluation compared microstructural heterogeneity and tested sensitivity to participant age and clinical group differences.The examined metrics were FA and IFW.Source: §1 · p. 4
• The framework was tested in two independent depression cohorts: younger participants with early-life depression and elderly participants with late-life depression.Both cohorts were compared with age-matched healthy controls; the early-life cohort included remitted and treatment-resistant participants.Source: §1 · p. 4

2.1.1 Participants and clinical assessment

• The study included an early-life cohort of 74 participants and a late-life cohort of 60 participants, comprising depressed patients and healthy controls.Depression severity was assessed using MADRS; the ELD cohort included remitted and treatment-resistant groups, while LLD controls were age-matched.Source: §2.1.1 · p. 4, §2.1 · p. 5

2 Methods

• The tractometry workflow links diffusion MRI preprocessing, tractography, multi-compartment modeling, metric projection, and atlas-based bundle-profile analysis.Figure 1 depicts subject-level processing, atlas construction and sub-bundle generation, and association of subject streamlines with atlas clusters.Source: Figure 1 · p. 5, §2.2 · pp. 6–7

2.1.2 MRI scan acquisition

• At both sites and in both samples, participants underwent 3T MRI, including a T1-weighted MPRAGE scan and multi-shell dMRI acquired with CUSP for 6.42 minutes, using b-values from 1000 to 3000 s mm−1.Source: §2.1.2 · p. 6

2.2.1 Preprocessing; 2.2.2 Tractography

• Preprocessing corrected motion, eddy currents, and distortion, while probabilistic iFOD2 whole-brain tractography targeted 1 million streamlines.The pipeline used reverse phase-encoded b0 images, nonlocal-means filtering, and MRtrix3 tractography constrained by the brain mask and streamline-length criteria.Source: §2.2.1 · p. 7, §2.2.2 · p. 7

2.2.3 Multi-compartment model

• Assuming Gaussian compartmental diffusion, the MCM comprises two isotropic compartments for free and restricted water plus 0–3 anisotropic compartments based on fixel counts, with IFW and trapped-water fraction (IR).FA is a direction-dependent metric, and each streamline point is associated with the anisotropic compartment.Source: §2.2.3 · p. 8

2.3 Tractometry atlas; 2.3.1 Streamline atlasing

• The tractometry reference was an HCP105 bundle atlas comprising anatomically defined bundles, used to guide bundle extraction, cluster geometry, and fiber alignment.The atlas is described as the geometric reference for extracting clusters and aligning individual fibers.Source: §2.3.1 · p. 8, §2.3 · p. 8

2.3.1 Streamline atlasing

• The streamline atlas was constructed from 105 HCP subjects; fiber bundles were semi-automatically obtained using fiber tracking and ROI filtering, then expert-curated, with 71 white matter bundles investigated.Source: §2.3.1 · p. 8

2.3 Tractometry atlas

• The HCP tensor atlas is built iteratively by registering and averaging tensor images, then transforming each subject’s 71 bundles into common space and merging streamlines to preserve shape variability.The atlasing uses diffeomorphic Stable Vector Fields and a log-Euclidean framework to compute average transformations.Source: §2.3 · p. 9

2.4 Subject-level bundle correspondence and tractometry

• The optimal sub-bundle partition is selected automatically by maximizing mean streamline silhouette over k ∈ {2, 3}, avoiding empirical distance-threshold tuning for each bundle.Initial QuickBundles compression uses a 5 mm threshold; Fréchet distances are computed between centroids, and original streamlines inherit their cluster labels.Source: §2.3.2 · p. 10, §2.4 · p. 11
• Atlas sub-bundles are mapped to subjects through bundle segmentation, streamline-to-cluster assignment, and point-level association, enabling measurements at anatomically consistent locations across individuals.Subject bundles are identified with BundleSeg using the HCP105 atlas, and streamlines are assigned to the nearest atlas cluster.Source: §2.4 · p. 11, §2.4.1 · p. 11, §2.4.2 · p. 12

2.4.1 Bundle segmentation

• BundleSeg uses the HCP105 bundle atlas as reference, aligns atlas FA maps to each subject, and applies distance-based matching with a fixed 8 mm radius to label bundles consistently.This produces a labeled set of streamlines for each bundle of interest.Source: §2.4.1 · p. 11

2.4.2 Streamline-level cluster association

• Each subject streamline is assigned to its nearest registered atlas cluster by point-wise Euclidean distance, comparing both streamline orientations and retaining the more similar configuration.Source: §2.4.2 · p. 12

2.4.3 Point-level cluster association

• Nearest-Euclidean point matching can err in curved or U-shaped bundles, while index-position matching can misrepresent anatomical correspondence across differing streamline lengths or local stretching.The cited RadTract alternative uses an SVM to learn anatomical partitions but is computationally demanding.Source: §2.4.3 · p. 12
• The association metric jointly penalizes spatial separation and mismatches in relative streamline position, using an automatically calibrated augmented space and k-d-tree search.The position penalty is scaled to the subject streamline’s arc length, requiring no tuning across fiber sizes or resampling resolutions.Source: §2.4.3 · p. 12, §2.5 · p. 13

2.5 Tractometry analysis

• For each cluster point, assigned streamline-point values were averaged to form subject-level profiles for FA and IFW, using 50 points per cluster and a shared valid segment.Points outside the longest contiguous segment with valid data for every subject were excluded from subsequent analyses.Source: §2.5 · pp. 13–14

2.5.1 Bundle-level statistics

• The sub-bundle approach was evaluated across bundles and microstructural metrics against the classical full-bundle method.The assessment included whether clustering captured within-bundle heterogeneity reflected in mean FA and IFW differences.Source: §2.5.1 · p. 14, Microstructural homogeneity · p. 14

Microstructural homogeneity

• Homogeneity was quantified using the standard deviation of microstructural values within clusters, comparing sub-bundle summaries with full-bundle variability.Analyses included mean sub-bundle SD and the SD of the most homogeneous sub-bundle, with Welch’s t-tests across bundles.Source: Microstructural homogeneity · p. 14

2.5.2 Along-tract statistics

• Point-wise Pearson age correlations were computed in healthy controls, while Welch’s t-tests compared clinical groups along bundles with cluster- and point-level family-wise error correction.Retained significant clusters had to include at least one point below the point-level corrected threshold.Source: §2.5.2 · p. 15

3 Results

• Among bundles split into two or three sub-bundles, within-cluster Fréchet distance was 36% or 29% lower than distance to the nearest neighboring sub-bundle, respectively.The resulting decomposition comprised 155 sub-bundles.Source: §3 · p. 15
• All ELD bundles had at least one sub-bundle associated with every subject, while 8 of 71 LLD bundles had missing subjects across sub-bundles and were excluded from group comparisons.Four bundles also lacked healthy controls in every sub-bundle and were excluded from age-sensitivity analyses.Source: §3 · pp. 15–16

3.1 Microstructural homogeneity

• Sub-bundle clusters were more homogeneous than full bundles (p < 0.01 to p < 0.001); significant homogeneity gains appeared in 42 ELD bundles for FA and 47 for IFW.For LLD, gains appeared in 48 bundles for FA and 34 for IFW; selecting the most homogeneous cluster reduced SD by a further 0.7% to 1.0% (p < 0.001).Source: §3.1 · p. 16, Figure 3 · p. 17

3.2 Age sensitivity

• In LLD controls, age correlated positively with IFW and negatively with FA; sub-bundles detected 11 significant bundles versus 5 with full-bundle analysis.The full-bundle analysis alone detected left FX for IFW and right ST PREM.Source: §3.2 · p. 16, Figure 4 · p. 18

3.3 Group comparison

• In Resistant vs. Remission, the cluster approach detected 10 significant bundles versus 5 with full-bundle analysis.The authors describe this as increased sensitivity to localized microstructural alterations.Source: §3.3 · p. 20

3.3.1 ELD

• For Resistant vs. Remission in ELD, the average FA difference increased from 1.8% ± 0.3 to 2.9% ± 0.6 with sub-bundle profiles.The sub-bundle and full-bundle analyses identified largely overlapping but method-specific significant bundles: 7 were unique to clustering and 6 to full bundles.Source: §3.3 · p. 20

3.3.2 LLD

• In LLD, sub-bundles identified 6 vs. 4 significant FA bundles and 23 vs. 15 IFW bundles, with mean differences of FA: −2.4% vs. −2.2% and IFW: 2.5% vs. 2.1%.The comparison is between sub-bundle and full-bundle analyses.Source: Table 2 · p. 20, §3.3.2 · p. 21
• Sub-bundle cluster 1 identified localized FA reductions in LLD in left ST–PREF and right ST–PAR that the full-bundle approach did not detect.IFW increases in LLD were also localized to specific sub-bundles.Source: §3.3.2 · p. 21

4 Discussion

• Fréchet distance measures maximum rather than average trajectory separation, while automatic quality-driven cluster-number selection avoids empirical geometric thresholds.Source: §4 · p. 21
• Sub-bundle clustering produced more homogeneous profiles than full-bundle tractometry, supporting the characterization of intra-bundle microstructural heterogeneity.Source: §4 · p. 21
• Age-related effects included increased IFW and decreased FA; sub-bundle analysis retained whole-bundle directions while detecting additional localized effects.Source: §4 · p. 23
• In LLD, gains were concentrated in fanning topographies, with thalamo-cortical and striato-cortical projections accounting for most IFW bundles detected only by sub-bundle analysis.Source: §4 · p. 23
• In ELD, alterations centered on deep callosal subdivisions and involved FA only, where parallel fibers offer less advantage for trajectory-divergence-based clustering.Source: §4 · p. 23
• Atlas quality and registration accuracy constrain cluster assignment and profile reliability; excluding incomplete sub-bundles balances samples but may introduce selection bias.Source: §4 · p. 24
• Fixed subdivision may suit compact tracts poorly; evaluation covers only FA and IFW, and the framework remains point-wise rather than fully three-dimensional.Source: §4 · pp. 24–25
• Declaring a bundle significant when any sub-bundle is significant slightly inflates bundle-level family-wise error; extending correction across clusters is proposed.Source: §4 · p. 25

5 Conclusion

• The framework decomposes bundles into geometrically coherent sub-bundles and selects their number by cluster separation without per-bundle parameter tuning.Source: §5 · p. 25
• Sub-bundles were more homogeneous than full bundles in FA and IFW and identified more age-correlated bundles across both cohorts and metrics.Source: §5 · p. 25
• New age-related detections arose in specific sub-bundles rather than whole tracts, indicating effects can be missed by single-profile summaries.Source: §5 · p. 25
• Group-discriminability gains were larger in late-life depression, where aging-related changes affect fanning projection tracts.Source: §5 · p. 25
• The gains were more modest in younger treatment-resistance comparisons, where alterations concentrate in deep, geometrically coherent tracts such as the corpus callosum.Source: §5 · p. 25
• Finer-scale tractometry and automatic clustering are presented as practical for group studies of subtle, spatially heterogeneous white-matter alterations.Source: §5 · pp. 25–26

Author contributions statement

• N.D. performed the statistics, interpreted results, and wrote the first draft.Source: Author contributions statement · p. 26
• J-C.R. and T.D. designed the study, performed data acquisition, and revised the manuscript.Source: Author contributions statement · p. 26
• M-L.P.M. found financial support and revised the manuscript; G.H.R. and J.C. designed the study, secured support, acquired data, supervised, interpreted results, and revised it.Source: Author contributions statement · p. 26

Ethics

• All methods were performed in accordance with the Declaration of Helsinki.Source: Ethics · p. 26
• The study was approved by the relevant institutional review boards under the listed study identifiers.Source: Ethics · p. 26
• All subjects provided informed, written consent before participation according to French legislation.Source: Ethics · p. 26

Data and Code Availability

• Original datasets are available from the corresponding author, subject to approval by the requesting researcher’s local ethics committee.Source: Data and Code Availability · p. 26
• Aggregated data and code used to generate the figures are publicly available at the project repository.Source: Data and Code Availability · p. 26
• Code used to process the original datasets is available at a separate repository.Source: Data and Code Availability · p. 26

• The study received support from the Fondation de l’Avenir, Fondation Planiol, and Fondation de France.Source: Funding · p. 26
• The Comité de recherche Clinique et d’Evaluation Thérapeutique at Rennes University Hospital also supported the study.Source: Funding · p. 26
• J-CR and ND received financial support from the Institut de Neuroscience Clinique de Rennes.Source: Funding · p. 26

• The authors report that AI was not used in writing the paper.Source: Declaration of the use of AI · p. 27

• Table S1 reports subject counts and mean age with standard deviation for healthy controls and the LLD group.Source: Table S1 · p. 34
• Table S2 assesses age comparability among healthy controls, remitted participants, and treatment-resistant participants in the ELD cohort.Source: Table S2 · p. 34
• Table S3 assesses age, MADRS, and depression-duration comparability between remitted and treatment-resistant ELD participants.Source: Table S3 · p. 34

S2 Point-level association

• Point-level association embeds streamline points using spatial coordinates and normalized position, then matches each to its nearest cluster point for anatomically corresponding metric aggregation.The composite distance balances spatial displacement and length mismatch.Source: Figure S1 · p. 35

S3 Incomplete sub-bundle associations

• Figure S2 presents missing associations by centroid and LLD group for excluded bundles without a complete sub-bundle, alongside color-coded sub-bundles and centroids.Source: Figure S2 · p. 36

S4 Bundle profiles

• Figures S3–S5 show group-mean along-tract profiles, shaded standard deviations, point-wise −10 log(p-values), and red multiple-comparison-significant clusters for ELD FA and LLD IFW comparisons.The captions identify ELD FA profiles in Figures S3 and S5, and LLD IFW profiles in Figure S4.Source: Figure S3 · p. 37, Figure S4 · p. 38, Figure S5 · p. 39