PaperlayerView exact revision on arXiv

Source-linked AI summary

Surface-volume self-supervised representation learning of brain MRI for genetic discovery

Tian Xia, Nuo Chen, Zihao Zhu, Huiwen Han, Ziqian Xie, Zhiwen Fan, Degui Zhi

arXiv:2610.02114v1cs.CV

TL;DR

Brain-imaging GWAS phenotypes have come from volumetric scans or cortical surface meshes, although these sources capture only part of heritable brain anatomy. MEVA combines volumetric MRI and cortical surface representations through residual fusion. It improves age and sex prediction and identifies more genome-wide significant loci than either single-modality representation.

  • Problem

    Brain-imaging genetic studies use volumetric or surface-derived phenotypes, while prior evidence indicates that these sources capture partly distinct heritable variation.

    Source: Introduction · p. 3

  • Method

    MEVA separately encodes volumetric MRI and cortical surface features, then combines the representations through residual fusion.

    Source: Introduction · p. 3

  • Results

    MEVA identified 188 genome-wide significant loci versus 153 volume-only and 40 mesh-only, with brain-age correlation 0.776 versus 0.768 and 0.671, and sex AUC 0.984 versus 0.976 and 0.886.

    Source: Discussion · p. 7

  • Takeaways & Limitations

    Residual fusion preserves genetic signal from the stronger volumetric stream while extending the representation with complementary mesh-derived information.

    Source: Discussion · p. 8

    The analyses do not yet characterize the biology underlying MEVA-specific loci; downstream genetic and explainability analyses are underway.

    Source: Discussion · p. 8

Abstract

from arXiv · show

Existing genome-wide association studies (GWAS) of brain imaging provide predefined or deep-learning-derived imaging phenotypes, yet these phenotypes come from either volumetric scans or cortical surface meshes, so each captures only part of the heritable variation in brain anatomy. Here we introduce MEVA (Mesh-Enhanced Volumetric Autoencoder), a self-supervised framework that encodes voxel-level image intensity together with cortical mesh geometry, including curvature and cortical thickness at each surface vertex, into one shared set of imaging features. Combining the mesh and volumetric inputs in MEVA yields modest performance gains in age and sex prediction over models that use either input alone. When these features serve as phenotypes for GWAS in the UK Biobank, they reveal more genome-wide significant loci than features learned from volumes alone or from meshes alone. These results suggest that adding cortical surface geometry to volumetric self-supervised learning captures additional heritable variation and so increases the number of loci detected.

Introduction

Brain-imaging genetic studies use volumetric or surface-derived phenotypes, and prior findings point to partly complementary heritable signals. The section introduces MEVA’s residual fusion of volumetric MRI and cortical mesh features and reports that mesh-only discoveries are weaker but partly distinct.

Introduction

  • Volumetric self-supervised representations support genetic discovery, but operate on 3D intensity arrays alone and leave folded cortical geometry unmodeled.Prior work includes autoencoder and vision-transformer approaches to MRI representations.Source: Introduction · p. 2
  • Mesh- and shape-based genetic studies report signal beyond volumetric traits, while no prior work had used self-supervision for mesh-based brain representations.Prior examples span cortical shape, thickness, surface area, and subcortical geometry.Source: Introduction · pp. 2–3
  • Volumetric and surface-derived genetic findings overlap only partly, consistent with geometry encoding variation partly independent of gross volumetric measures.Most cited comparisons rely on catalog overlap rather than formal colocalization or fine-mapping, leaving the relationship between signal sources incompletely resolved.Source: Introduction · p. 3
  • MEVA separately encodes volumetric MRI and cortical surface features, then combines their representations through residual fusion.The mesh features are derived from FreeSurfer-reconstructed surfaces.Source: Introduction · p. 3
  • Mesh-only GWAS identified 40 significant loci versus 153 for the volumetric representation, with only partial overlap; MEVA is reported to exceed either single-modality baseline.The mesh stream is described as carrying weaker overall genetic signal while capturing information only partly represented by the volumetric stream.Source: Introduction · p. 4
  • Residual fusion keeps the volumetric representation primary while adding mesh information, addressing the risk that unconstrained fusion could suppress mesh signal or disrupt stronger volumetric signal.When mesh contribution is small, the fused representation falls back toward the volumetric one.Source: Introduction · p. 4

Results

Mesh-only phenotypes captured age and sex information, while MEVA modestly improved those predictions over single-modality models. In genetic discovery, MEVA yielded the most loci and the highest summed SNP heritability among the compared representations.

Phenotypic information in cortical mesh representations

  • Mesh-only phenotypes built from vertex coordinates, cortical thickness, and curvature predicted age at r = 0.671 and sex at AUC = 0.886.Cortical thickness is the distance between the pial and white-matter surfaces; curvature ranges from sulcal concavity to gyral convexity.Source: Phenotypic information in cortical… · p. 4, Figure 1 · p. 5

Integrating cortical geometry and volumetric intensity

  • MEVA phenotypes predicted age at r = 0.776 versus r = 0.768 for volume-only and r = 0.671 for mesh-only, and sex at AUC = 0.984 versus 0.976 and 0.886, respectively.All representations had 128 phenotypes and were evaluated in the same 35,298 held-out UK Biobank participants.Source: Phenotypic information in cortical… · p. 5, Integrating cortical geometry and… · p. 6
  • MEVA combines a volumetric T1-weighted MRI encoding with a pretrained, frozen mesh encoding through attention and a residual added to the volume representation.The resulting fused representation has 128 dimensions and is trained with T1 reconstruction and mesh-embedding prediction losses.Source: Figure 2 · p. 6

Genetic discovery from multimodal representations

  • Forty-two loci were identified only by MEVA, with loci treated as overlapping when their intervals padded by 125 kb on each side overlapped on the same chromosome.The MEVA minimum-P GWAS used N = 35,298 participants.Source: Genetic discovery from multimodal… · p. 7, Figure 3 · p. 7
  • MEVA identified 188 loci versus 153 for volume-only and 40 for mesh-only, and had summed SNP heritability 36.7 versus 33.7 and 10.5, respectively.The three representations were compared using the same participants, covariates, and association pipeline.Source: Genetic discovery from multimodal… · p. 6, Figure 3 · p. 7

Discussion

MEVA combines a volumetric MRI representation with cortical mesh features and yields more informative phenotypes and greater locus discovery than either modality alone. The discussion also identifies untested fusion strategies and notes that the biology of MEVA-specific loci remains uncharacterized.

Discussion

  • 188 genome-wide significant loci were identified by MEVA, versus 153 for volume-only and 40 for mesh-only representations; 42 were found by neither unimodal representation.MEVA also achieved the highest summed SNP heritability, 36.7 compared with 33.7 and 10.5.Source: Discussion · p. 7
  • Brain-age prediction reached Pearson correlation 0.776 with MEVA, versus 0.768 for volume-only and 0.671 for mesh-only; sex-classification AUC was 0.984, versus 0.976 and 0.886.Source: Discussion · p. 7
  • Cortical-thickness variance explained was ΔR2 = 0.375 for MEVA, versus 0.091 for the volume-only model.The authors interpret the additional loci as a gain in captured heritable variation rather than redistribution across latent dimensions.Source: Discussion · p. 7
  • Residual fusion combines a strong primary modality with a weaker but complementary one, preserving volumetric genetic signal while extending it with mesh-derived information.Source: Discussion · p. 8
  • Alternative designs include contrastive alignment, shared-versus-specific latent allocation, and local cross-attention; contrastive objectives may attenuate mesh-specific signal, while cross-attention raises computational and memory demands.The authors identify systematic comparison of these strategies under the same GWAS and evaluation framework as future work.Source: Discussion · p. 8
  • The analyses do not yet characterize the biology underlying MEVA-specific loci; downstream genetic and explainability analyses are underway.Source: Discussion · p. 8

Methods

The study extracted phenotypes for a UK Biobank genetic-analysis cohort and used quality-controlled T1-weighted MRI with standard preprocessing. Model training used a separate subset of scans, with usable cortical surfaces required for mesh-based models.

UK Biobank participants

  • The genetic-analysis cohort comprised 35,298 participants with T1-weighted MRI, cortical surfaces, and imputed genotypes; mean age at imaging was 67.8 years.These participants were not used for model training.Source: UK Biobank participants · p. 9

MRI acquisition and volumetric preprocessing

  • UK Biobank supplied brain-extracted, bias-field-corrected T1 scans acquired at 1-mm isotropic resolution on Siemens Skyra 3T scanners; affine registration aligned images to MNI152 space while preserving subject-specific structural deformations.Images were intensity-normalized per scan and zero-padded to 182 × 224 × 182 voxels.Source: MRI acquisition and volumetric… · p. 9
  • The filter retained scans below the 95th percentile on both inverted contrast-to-noise ratio and T2 FLAIR–T1 discrepancy metrics, yielding 6,130 scans; 6,067 participants had usable cortical surfaces for mesh autoencoder and MEVA training.The volumetric autoencoder used 4,597 training and 1,533 validation scans.Source: MRI acquisition and volumetric… · pp. 9–10

Cortical surface preparation

Cortical surfaces were represented as vertex-wise geometry and morphometry, and a masked graph autoencoder learned a mesh embedding. MEVA fused this embedding with the volumetric representation through a residual contribution.

Cortical surface preparation

  • Each fsaverage4 hemisphere has 2,562 vertices, carrying mid-thickness coordinates, cortical thickness, and curvature as five channels.Coordinates were centered and scaled using training-set estimates, while thickness and curvature were standardized using training-set statistics.Source: Cortical surface preparation · p. 10

Mesh autoencoder

  • The mesh-only encoder is a masked graph autoencoder that reconstructs five vertex channels and averages vertex features within each hemisphere before producing a 128-dimensional embedding.It uses shared weights across hemispheres, masks 60% of vertices, and was trained on 4,853 participants with 607 each for validation and held-out testing.Source: Mesh autoencoder · p. 10

MEVA fusion and training

  • MEVA combines the pretrained volumetric autoencoder with a frozen mesh encoder, and a fusion module computes a residual from the volumetric and mesh embeddings.The residual is added to the volumetric representation, which is passed to the convolutional decoder for T1 reconstruction.Source: MEVA fusion and training · p. 10
  • Fusion training proceeded in two stages: a residual multilayer perceptron with masked-T1 reconstruction and InfoNCE, followed by a two-token transformer with mesh distillation.Surface information enters through the mesh embedding and distillation loss rather than vertex-feature reconstruction; added modules total about 0.18 million parameters.Source: MEVA fusion and training · pp. 10–11

Volumetric autoencoder baselines and training

The volume-only baseline uses a pretrained CNN autoencoder, which also initializes MEVA’s volumetric encoder. Encoders were frozen to extract comparable 128-dimensional phenotypes for genetic analyses.

Volumetric autoencoder baselines and training

  • The volume-only representation is the pretrained CNN autoencoder with 138.1 million parameters, trained for 40 epochs and used without further training.The same model initializes the MEVA volume encoder.Source: Volumetric autoencoder baselines and… · p. 11
  • The volumetric autoencoder has five encoding and five decoding CNN blocks, 3 × 3 × 3 kernels, 2 × 2 × 2 max pooling, and a 128-dimensional fully connected layer.Source: Volumetric autoencoder baselines and… · p. 11

Embedding extraction

  • Frozen volume-only, mesh-only, and MEVA encoders produced 35,298 × 128 phenotype matrices from the same genetic-analysis participants.The MEVA encoder takes both the T1 image and the participant’s mesh embedding.Source: Embedding extraction · p. 11

Phenotypic prediction and statistical analysis

This section describes phenotype prediction and cortical-information metrics, genotype quality control, and the procedures for GWAS and heritability estimation.

Phenotypic prediction and statistical analysis

  • Age prediction used ridge-regression correlation and mean absolute error, sex prediction used class-balanced logistic-regression AUC, and cortical information was average incremental variance explained across regional measures.Cortical measures included thickness, surface area, and grey-matter volume in 68 Desikan–Killiany regions.Source: Phenotypic prediction and statistical… · p. 11
  • Confidence intervals used 500 participant bootstrap resamples; for ΔR², both models were refitted per resample and intervals were estimate ±1.96 bootstrap standard errors.Source: Phenotypic prediction and statistical… · pp. 11–12

Genotype quality control

  • Genotype quality control removed duplicated variants and excluded individuals whose genetically inferred sex disagreed with self-reported sex.Source: Genotype quality control · p. 12

Genome-wide association analysis

  • 8,931,083 variants were tested across 128 phenotypes in 35,298 participants using fastGWA, with genome-wide significance set at P < 5 × 10^-8/128.Models included age, age², sex, their interactions, and genetic principal components.Source: Genome-wide association analysis · p. 12
  • Total SNP heritability was estimated by PCA-decorrelating phenotypes and summing single-phenotype Haseman–Elston estimates.The stated total-heritability definition is tr(P^-1G).Source: Genome-wide association analysis · p. 12

Genomic locus annotation

The section describes how genomic loci were assigned, how overlap was assessed, and the study's ethics approvals and consent basis.

Genomic locus annotation

  • Lead SNPs were identified using linkage disequilibrium (r² ≤ 0.1) and physical proximity (<250 kb), then assigned to non-overlapping loci represented by each locus's lowest-p-value SNP.Source: Genomic locus annotation · p. 12

Locus overlap analysis

  • For overlap analysis, each locus was extended 125 kb on each side, and chromosome-specific interval trees were queried for intersecting regions.Source: Locus overlap analysis · p. 12

Ethics approval and consent

  • The study was approved by the UTHealth Houston human-subjects committee, and UK Biobank participants had consented to approved research use.UK Biobank data were accessed under approved project 24247.Source: Ethics approval and consent · p. 12

Data availability

The paper describes routes to request its analysis outputs and unrestricted supporting data, as well as access to UK Biobank individual data.

Data availability

  • Self-supervised imaging phenotypes and GWAS summary statistics are available upon reasonable request after sensitive participant information is removed.Source: Data availability · p. 12
  • Individual UK Biobank data can be requested with proper registration.Source: Data availability · pp. 12–13
  • Unrestricted data supporting the findings are available from the corresponding author upon request.Source: Data availability · p. 13

Supplementary Information

The supplementary material includes a notation table, a table on model capacity and training, and descriptions of the model representations and components.

Supplementary Information

Loading 2610.02114v1…