Marvin Petersen

📢 DGKN 2026: Congress for Clinical Neurosciences with Continuing Education Academy
📅on 25 February 2026, Prof. Dr. Thomas Wolfers
will present a talk titled: “𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗮𝗻𝗱 𝗕𝗶𝗴 𝗗𝗮𝘁𝗮 𝘁𝗼 𝗣𝗮𝗿𝘀𝗲 𝗜𝗻𝗱𝗶𝘃𝗶𝗱𝘂𝗮𝗹 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀”
@thomaswolfers.bsky.social #DGKN2026 #ClinicalNeuroscience #BigData #ML

We don't think it invalidates the approach.

We looked into this in our data and replicated the key result showing convergence to non-specific connectome structure. However, our results also suggest there is a practical fix that is already implemented in some LNM studies.

bsky.app/profile/pete...

Many thanks to Geert Jan Biessels, Kaustubh Patil, @sbe.bsky.social, and all colleagues in the Meta VCI Map Consortium (names in the supplement)!

Code is available on GitHub: github.com/umcu-VCI-gro...

Domain-level lesion network maps are on OSF: osf.io/puq36

The takeaway: LNM is still a valuable tool when the statistics are handled correctly. Symptom-label permutation provides a practical fix for the convergence problem, indicating that patient-level connectivity maps do carry genuine, symptom-specific signals that established techniques can miss.

Enhancing cognitive performance prediction by white matter hyperintensity connectivity assessment Petersen et al. show that assessing the brain circuitry affected by cerebrovascular white matter lesions improves the accuracy with which cognitive perform

Jointly, these results align with our work showing that LNM significantly improves individual-level symptom prediction compared to pure lesion location – something that would not be possible if patient-level lesion connectivity maps would only carry non-specific information.

doi.org/10.1093/brai...

Comparison of label permutation-based lesion network mapping and voxel-based lesion-symptom mapping (VLSM). For each cognitive domain, lesion network mapping (LNM) results thresholded at pFDR < 0.05 (red) are overlaid with VLSM maps (blue). Dice coefficients reported in each panel quantify the overlap between methods after significance thresholding for each outcome. Permutation-based LNM maps showed only partial overlap with VLSM maps for the same outcomes, suggesting that even under label permutation testing, LNM captures effects that are not simply reducible to focal lesion-symptom associations.

Importantly, the label permutation approach did not reduce LNM to focal lesion effects: permutation-based LNM maps only partially overlapped with voxel-based lesion-symptom mapping results for the same outcomes.

Comparison of parametric and label-permutation inference in symptom-linked lesion network mapping. For each cognitive domain, group-level lesion network maps thresholded at pFDR < 0.05 are shown for parametric voxelwise inference (blue) overlaid with label permutation maps (red; appearing darkblue where overlapping; note that no voxels light up red, because all label permutation map voxels overlap with parametric voxels). Dice coefficients report the overlap between parametric and label permutation maps after significance thresholding for each outcome.

This plot demonstrates differences between parametric (blue) and permutation-based maps (red).

Permutation-based inference reduces convergence toward degree-like structure and decreases cross-domain similarity. (a) Symptom-linked lesion network maps derived using label permutation, thresholded at pFDR < 0.05. (b) Pairwise comparisons of label-permutation lesion network maps. Upper triangle: similarity quantified as Dice overlap of significance-thresholded maps. The first row/column compares each map with the normative connectome degree map (degree defined as the row-sum of the normative FC matrix; binary degree mask thresholded at the 95th percentile). Lower triangle: pairwise overlaps of significance-thresholded maps; maps corresponding to columns are shown in blue and rows in red. Diagonal: volumes of thresholded masks (mm³).

Under permutation-based inference, the story changed. Maps became more biologically plausible (e.g., the language network lateralized), cross-domain similarity dropped, and degree-like overlap decreased. Reflecting this stricter null, visuospatial memory lost significance entirely.

How it works: Shuffling symptom labels reproduces the convergence in every permutation. Thus, the artifact becomes part of the null distribution. Under this stricter null, only signals exceeding the artifact remain significant, treating the rest as nonspecific background.

We implemented symptom-label permutation with FSL PALM which just takes one line of code. web.mit.edu/fsl_v5.0.10/...

Mapping Lesion-Related Epilepsy to a Human Brain Network This case-control study evaluates the locations of epilepsy-associated lesions with respect to specific brain regions and networks.

Next, we applied symptom-label permutation, a standard nonparametric strategy in neuroimaging that has also been used in other LNM studies beyond our own (e.g., https:// doi.org/10.1001/jamaneurol.2023.1988).

Parametric symptom-linked lesion network mapping yields highly similar maps across cognitive domains. (a) Symptom-linked lesion network maps derived from parametric statistics, thresholded at pFDR < 0.05. (b) Pairwise comparisons of symptom-linked maps. Upper triangle: similarity quantified as Dice overlap of thresholded maps and Pearson correlation (r) of voxelwise T-statistics. The first row/column compares each symptom-linked map with the normative connectome degree map (degree defined as the row-sum of the normative FC matrix; binary degree mask thresholded at the 95th percentile). Lower triangle: voxelwise scatterplots of T-statistics. Diagonal: volumes of thresholded masks (mm³).

First, we reproduced the “standard” symptom-linked LNM workflow with parametric voxelwise inference as in the critique paper.

As predicted, we observed convergence under this approach. Distinct deficits were associated with nearly identical networks which aligned with the normative degree map.

We tested this formally in multicenter post-stroke cognitive impairment data from the Meta VCI Map Consortium (12 cohorts; n=2,950; metavcimap.org). We statistically compared lesion connectivity maps of cognitively impaired vs. unimpaired stroke patients across 6 distinct cognitive domains.

Enhancing cognitive performance prediction by white matter hyperintensity connectivity assessment Petersen et al. show that assessing the brain circuitry affected by cerebrovascular white matter lesions improves the accuracy with which cognitive perform

The paper prompted us to revisit our own earlier work, which used symptom-label permutation instead of parametric statistics as commonly used in LNM. We realized this permutation approach provides a straightforward way to address the highlighted convergence problem.

doi.org/10.1093/brai...

As group-level aggregation repeatedly samples the same connectome matrix, different lesion sets produce network maps of implausibly high similarity. Consequently, the authors question whether LNM can identify symptom-specific networks.

We agree that these findings necessitate a reappraisal of LNM.

LNM links focal brain lesions to distributed neural circuits using a normative functional connectome. However, van den Heuvel et al. showed that common LNM procedures introduce a bias that causes maps to converge on nonspecific properties of the connectome, such as node degree.

Investigating the methodological foundation of lesion network mapping - Nature Neuroscience The lesion network mapping method links diverse brain lesions to similar functional brain networks, reflecting general brain organization rather than disorder-specific circuits.

In this work, we address the recent methodological critique on lesion network mapping (LNM) by Martijn van den Heuvel and colleagues (doi.org/10.1038/s415...). We replicate their results in a multicentric sample of 2,950 stroke patients and propose a practical fix to the identified problem!

Revisiting the methodological foundation of lesion network mapping Lesion network mapping (LNM) links focal brain lesions to distributed neural circuits by projecting lesion locations through a normative functional connectome. van den Heuvel and colleagues recently s...

New preprint: Revisiting the methodological foundation of lesion network mapping.

doi.org/10.64898/202...

A thread.

"You see what you look for, and you look for what you know."

We are proud to release FOCUS: a 100 μm deformable human subcortical atlas precisely registered to MNI space.

www.biorxiv.org/content/10.6...

Nine subcortical/cerebellar atlases included in the subcortex_visualization Python package (and subcortexVisualizationR package in R). The atlases are depicted in two-dimensional vector graphic format.

The extended version of my thesis procrastination project/subcortex visualization package is out now in both Python and R, now that I’ve graduated 🤠 This figure shows the 9 atlases included (and counting)!

Preprint: www.biorxiv.org/content/10.6...
Website: anniegbryant.github.io/subcortex_vi...

netneurotools: a trainee-oriented approach to network neuroscience | doi.org/10.1101/2025...

Our lab’s internal toolkit for accomplishing everyday tasks in brain imaging ⤵️

🍤 hippocampus = the brain's memory machine 🍤

unique cytoarchitecture, highly connected with multiple networks, & very variable across individuals

video shows hippocampal shape variations across 250-ish young HCP adults

study hippocampus with open tools like hippomaps doi.org/10.1038/s415...

1/
Our new paper is out in Imaging Neuroscience! 🎉
We validated in-vivo MRI–based axon radius mapping by comparison to large-scale human brain histology (> 46 M axons & 35 CC-ROIs).
Title: “MRI-scale histology validates spatial sensitivity of in-vivo MRI-based axon radius estimation”

Diffusion MRI Processing in the HEALthy Brain and Child Development Study: Innovations and Applications The landmark ongoing HEALthy Brain and Cognitive Development (HBCD) study will longitudinally chart brain development in a large sample (projected n=7,200) of infants through age 10 years with multimo...

After years of development and testing, we are happy to present our work in "Diffusion MRI Processing in the HEALthy Brain and Child Development Study: Innovations and Applications"! www.biorxiv.org/content/10.1.... A thread:

This work was made possible through the contributions of an excellent team of collaborators @csi-lab.bsky.social @felenae.bsky.social

Together, these results suggest that gray matter features - particularly those reflecting microstructural integrity - may complement the existing pool of imaging biomarkers for early CSVD-related cognitive decline.

Our second key finding: Gray matter integrity measures in these regions were significantly associated with lower general cognitive ability after adjustment for age, sex, and education.

Our first key finding: Higher CSVD burden was linked to widespread, regionally specific gray matter abnormalities. These included altered diffusivity, tissue integrity, and thickness in the cingulate, insular, and temporal cortices, as well as the hippocampus.

We analyzed neuroimaging and cognitive data from 2,603 participants of the population-based Hamburg City Health Study. Using multi-modal MRI we derived a composite CSVD burden score and assessed regional gray matter micro- and macrostructure.