Organizations: D. Bradley McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, US. · Department of Biomedical Informatics and Data Science, Yale School of Medicine, New Haven, US. · Department of Electrical and Computer Engineering, Texas A&M University, College Station, US. · Department of Psychiatry and Behavioral Sciences, Stanford University School of Medicine, Stanford, CA, USA.
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.
Figures & tables
Figure 1: Mesh-only autoencoder. A masked graph autoencoder encodes the fsaverage4 cortical mesh of each hemisphere, reconstructs the masked vertex features and reads out the embedding zmesh . Each vertex carries its mid-thickness x , y and z coordinates, cortical thickness (the distance between the pial and white-matter surfaces) and curvature, from concave (sulcal fundi) to convex (gyral crowns).
Prediction
Cortical ΔR2↑
Genetics
Representation
Age Pearson r ↑
Sex AUC ↑
Thickness
Area
Volume
Loci ↑
∑hSNP2↑
Volume-only
0.768
0.976
0.091
0.258
0.209
153
33.7
Mesh-only
0.671
0.886
0.475
0.284
0.294
40
10.5
MEVA
0.776
0.984
0.375
0.277
0.266
188
36.7
Table 1: Comparison of the learned representations. 0 0 footnotetext: All representations have 128 phenotypes and are evaluated in the same 35,298 held-out UK Biobank participants; bold marks the best value in each column. Age mean Pearson r and sex AUC are from five-fold cross-validation. Cortical information is the variance of FreeSurfer regional thickness, surface area and grey-matter volume explained beyond covariates, averaged over the 68 Desikan–Killiany regions. Bootstrap 95% confidence intervals lie within ±0.006 of every sex AUC and ΔR2 value; Loci are from fastGWA analysis of 128 phenotypes, followed by the minimal p value of aggregating multivariate summary statistics and FUMA for loci aggregation; ∑hSNP2 is the summed SNP heritability from the sum-h2 pipeline (Methods).
Figure 2: MEVA architecture, training objective and phenotype extraction. a , A T1-weighted MRI volume and its FreeSurfer-reconstructed cortical surface are used as inputs. b , A 3D CNN encodes the volume into zvol , while a pretrained, frozen mesh autoencoder encodes the surface into zmesh . A self-attention block jointly processes the embeddings, and its output is projected to a residual δ that is added to zvol to obtain zfused∈R128 . Training uses T1 reconstruction and mesh-embedding prediction losses, L=LT1+λLmesh ( λ=10 ). c , The trained encoder generates 128 imaging phenotypes from each participant. These phenotypes are used for genetic discovery and anatomical evaluation, including age, sex and cortical ΔR2 prediction.
Figure 3: Genetic associations across imaging representations. a , Overlap of genome-wide significant loci identified by the volume-only (153 loci), mesh-only (40 loci), and MEVA (188 loci) representations; loci were merged across models when their boundaries, padded by 125 kb on each side, overlapped. Forty-two loci were identified only by MEVA. b , Manhattan plot of the MEVA minimum-P GWAS ( N=35,298 ). SNPs within the 42 MEVA-specific loci are highlighted in yellow.
Symbol
Meaning
x , x^
Input T1-weighted volume and its reconstruction
zvol∈R128
Embedding of the T1 volume from the volumetric CNN encoder; the volume-only phenotypes
zmesh∈R128
Embedding of the cortical surface from the frozen mesh encoder; the mesh-only phenotypes
zmesh′
zmesh after the learned adapter
δ∈R128
Surface-informed residual computed by the fusion module
zfused=zvol+δ
MEVA phenotypes
Table S1: Notation.
Module
Parameters
Training
CNN autoencoder
138.1M
40 epochs (pretraining)
Mesh autoencoder
0.03M
epoch 98 of 100
MEVA fusion modules
0.18M
4 + 6 epochs (two stages)
Table S2: Model capacity and training. 0 0 footnotetext: The CNN autoencoder gives the volume-only phenotypes and initializes the MEVA volume encoder, which is fine-tuned during fusion; it has 69.6M encoder and 68.5M decoder parameters. The mesh autoencoder gives the mesh-only phenotypes and is frozen in MEVA. The MEVA fusion modules are the mesh adapter, the fusion block and the distillation head.
Artificial Intelligence in Biomedical Imaging Lab, University of Pennsylvania · Department of Electrical and Systems Engineering, University of Pennsylvania · Florey Institute of Neuroscience and Mental Health, University of Melbourne +16