Can representation learning of cortical folding reveal lasting neurodevelopmental signatures?

Folding is laid down early and then stays put

The paper's premise is a timing argument. The human cortex folds mainly during the second half of gestation and reaches near-adult folding levels shortly after birth, so the shape and topology of folds are established early and remain remarkably stable thereafter. Scalar descriptors such as sulcal depth or the gyrification index behave differently: they change across development and aging. That contrast is what makes folding attractive for probing early neurodevelopmental mechanisms rather than later disease consequences.

The authors also argue that folding variability is not random. Before folding, spatial gradients of gene expression in the developing cortex closely match the future layout of sulci and gyri, and folding patterns partially predict the underlying cytoarchitecture. Prior work has linked folding to behavioral and clinical phenotypes, but most of it relies on manual annotations or a small number of handcrafted descriptors, which limits how systematically folding variability can be explored at scale. Champollion is the authors' answer to that scale problem.

Champollion learns one small region at a time

Champollion is a self-supervised learning framework, meaning it learns features directly from imaging data without predefined anatomical measurements. Its distinctive choice is scale: instead of one global brain representation, it learns a low-dimensional representation for each region of interest (ROI), typically centered on one sulcus or a few neighboring sulci. The authors give two reasons. Analyses in the folding community are usually conducted on predefined regions, and local latents are inherently interpretable because each variable is tied to an anatomical region, enabling spatial maps of association and visual characterization of individual sulcus patterns.

The preprocessing is designed to preserve folds rather than smooth them away. Standard pipelines such as voxel-based morphometry or surface-based registration use non-linear alignment, which the authors say inevitably distorts cortical folds and discards exactly the information they want. Instead they skeletonize the cortex into one-voxel-wide traces of the sulcal medial axis, which removes fold-opening information that aging affects; affinely align the skeletons to the MNI template at 2 mm isotropic resolution; and tile the skeleton into 56 overlapping ROIs, 28 per hemisphere, defined from an independent dataset of 62 hand-labeled subjects with a 10 mm dilation. Left and right hemispheres are treated separately because sulcal regions are not guaranteed to be bilaterally symmetric.

For the learning objective, the authors adopt the Barlow Twins loss rather than contrastive or reconstruction alternatives. They exclude transformer-based methods because they want lightweight convolutional local encoders, exclude SimCLR because negative pairs are ill-defined without clear class boundaries, and prefer Barlow Twins over BYOL and SimSiam because it explicitly regularizes feature covariance and is theoretically less prone to dimensional collapse. The redundancy reduction parameter is set to 8×10⁻².

Generic foundation models largely miss folding variability

The central comparison is a benchmark against three families of alternatives: handcrafted morphometric descriptors, previous self-supervised approaches, and a broad range of recent vision foundation models trained on either natural images or brain images. The paper reports that Champollion consistently outperforms all of them. The authors read this as evidence that current foundation models largely overlook cortical folding variability.

Their proposed explanation is architectural. Champollion's local formulation revealed a high degree of independence between folding patterns across cortical regions in the general population. The authors suggest that such localized information may be overlooked by neuroimaging models optimized to capture broad anatomical variability, which would partly explain the foundation-model gap. This is the authors' interpretation of the benchmark result, not a separately tested mechanism.

The optimization tasks themselves are worth noting because they define what "accurately captures known folding patterns" means here. The paper describes tasks built on known folding patterns across cortical regions, including regression of six continuous shape descriptors of the central sulcus in 880 HCP subjects derived via Isomap, and classification of anterior cingulate cortex patterns. The authors acknowledge that the Isomap descriptors are not ground truth, since they come from a computational method, but considered it reasonable to seek to capture as much of that information as possible.

Latent proximity tracks heritability better than classical descriptors

Beyond matching known patterns, the paper tests whether the representations carry biological information that classical morphometry does not. Two results support this. First, the ability to match monozygotic twins based on latent proximity provides evidence that the learned features reflect heritable aspects of cortical folding. Second, Champollion's representations exhibit substantially stronger genetic associations than the classical morphometric descriptors of folding.

The authors use this to argue that folding patterns are genetically grounded and that the oversight of folding in generic neuroimaging representations is a missed opportunity. The claim is comparative: the same underlying biology is better recovered by these learned local representations than by conventional descriptors. The paper does not report a specific effect size for the genetic comparison in the extracted text, so the strength of "substantially stronger" should be read as the authors' characterization rather than a number.

Three exploratory tasks, three localized folding signatures

The clinical analyses are framed by the authors as exploratory and of increasing difficulty. All classifications use linear probing of Champollion's representations after site harmonization with ComBat and residualization on sex, age, and socio-economic status when available. Performance is measured by ROC-AUC, significance is assessed with a permutation test using the ter Braak scheme, and the significance threshold is Bonferroni-corrected to 0.05/56, which sets 11,200 permutation tests per ROI. For each task, the best-performing ROI is then interpreted by latent traversal along the classification direction, with the direction constrained to explain at least 2% of global variance; the authors report this constraint barely affects classifier performance, under 2% on all investigated tasks.

Incomplete hippocampal inversion used the QTIM dataset with a single site, comparing pronounced IHI against absence (N=928, 8% pronounced IHI), with twins kept in the same fold. The identified signature was a torque reduction of the left collateral sulcus. The authors note this is anatomically consistent with the hippocampus location and that prior work reported no association between left IHI and manually defined sulcal variants in the collateral region based mainly on interruption patterns, which aligns with a previously unreported torsion rather than an interruption.

Great prematurity used the ABCD dataset, comparing very preterms born at 28–32 weeks post-menstrual age against full terms born after 37 weeks (N=8,846, 1% preterms). The signature was a flattening of the anterior right superior temporal sulcus. The authors connect this to prior work implicating the temporal lobe and to evidence that superior temporal sulcus depth at term-equivalent age correlates with gestational age at birth, with stronger effects in the right hemisphere.

Maternal smoking during pregnancy was fitted on UK Biobank using binary labels reported by the offspring, with no site harmonization because the three sites shared the same scanner and protocol, and then replicated in ABCD, which carries quantitative categories of average cigarettes per day. The signature was a global reorientation of folds in the left occipital region. The authors state they found no prior reports of associations between maternal smoking and sulcal patterns in the occipital region; existing studies mainly report weak effects on cortical thickness. They explicitly caution that these analyses illustrate a framework for hypothesis generation rather than establishing definitive biological associations.

The smoking signature replicates across two very different cohorts

The strongest evidence for the "long lasting" part of the title comes from the maternal smoking analysis. The pattern was replicated across UK Biobank, aged 44–82 years, and ABCD, aged 9–11 years. The authors read this as suggesting the detected signal reflects a stable, long-term signature. They add that successful transfer of classifiers across datasets with markedly different demographics indicates Champollion is robust to site and population variability.

The occipital finding is notable for a second reason the authors raise: the region is highly variable and lacks a standardized sulcal nomenclature in the anatomical literature. Finding a coherent pattern there is presented as evidence that the method can surface structure in regions where manual annotation is unreliable. The paper also notes that for sulci where automated labeling is less robust, label-agnostic methods such as Champollion may be better suited to capture continuous shape patterns than Isomap-based descriptors.

Hypothesis generation, not confirmed biology

Several limits qualify how far these results travel. The ROI atlas is one possible tiling rather than a unique optimal decomposition, so the localization of any signature depends on that choice. The identified patterns may correlate with other anatomical factors such as local brain volumes and global shape; the authors cite prior work suggesting these are not the primary determinants of folding variability but say further analysis is required to determine their specific contribution. Combining local and global representations may also reveal coordinated patterns lying in low-variance modes, which the current local design would not capture.

The clinical analyses are explicitly exploratory. The authors describe them as illustrating a general framework for hypothesis generation rather than establishing definitive biological associations, and the maternal smoking occipital pattern in particular has no prior report they could locate. The genetic comparison is characterized qualitatively as substantially stronger than classical descriptors rather than quantified in the extracted text.

The paper's own forward-looking use is anomaly detection derived from this model of typical variability, with potential applications to developmental disorders such as epilepsy, psychiatry, and rare genetic syndromes. That is a stated direction, not a demonstrated result.

The practical case rests on standard MRI and compact features

The authors make a practical argument that matters for anyone choosing a representation for folding analysis. Champollion relies on standard, widely available structural MRI, requires only moderate spatial resolution at 2 mm isotropic, and produces compact, region-specific representations that can be analyzed with simple statistical models such as linear probing. That combination is what makes the benchmark comparison consequential: the alternative is not a different deep model but often a handcrafted descriptor or a general foundation model that, on this evidence, captures less folding variability.

The priority the paper establishes is therefore specific. When the question concerns early, stable folding shape rather than scalar morphometry that drifts with age, the evidence favors a locally trained self-supervised representation over generic neuroimaging or natural-image foundation models, and favors interpreting results region by region rather than through a single whole-brain embedding. The maternal smoking replication across cohorts aged 9–11 and 44–82 is the clearest demonstration that a folding signature found this way can persist, while the IHI and prematurity findings remain hypothesis-generating leads tied to specific sulci.

Sources used in this answer