Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability
Gil Sasson, Zachary Levine, Smadar Shilo, Sarah Kohn, Guy Lutsker, Anastasia Godneva, Adam Gabet, David Krongauz, Adina Weinberger, Yann LeCun, Randall Balestriero, Eran Segal
cs.CV, q-bio.QM
2026-08-03
A 21.7M-parameter JEPA model trained on 11,540 unlabeled DXA scans beats scanner readouts and DINOv3 at predicting incident disease, biological age, and mortality risk.
DXA scans are among the most routinely acquired medical images: cheap, low-radiation, covering bone density and whole-body fat and muscle distribution in a single shot. Clinical practice keeps a handful of scalar readouts from each scan and discards nearly all spatial structure. Prior deep learning on DXA was mostly supervised, requiring fresh clinical labels for every new target, and the standard workaround of reusing frozen features from general-purpose models like DINOv2/v3 fits poorly: those models are tuned to the colors and edges of natural photos, while chronic disease shows up as body-wide morphological change. The bottleneck is data. DXA datasets are counted in the thousands, several orders of magnitude short of the millions of images usually deemed necessary to train a foundation model from scratch.
LeDXA trains from scratch instead. The backbone is a ViT-Small/16 (21.7M parameters, 384-dim output) trained with the LeJEPA framework on 11,540 unlabeled whole-body scans from the Human Phenotype Project, an Israeli deep-phenotyping cohort of 8,820 adults, with no ImageNet pretraining. JEPA predicts target representations in latent space rather than reconstructing pixels, so the model cannot get away with memorizing pixel noise and has to learn anatomical shape. Each scan is split into a bone-modality and a tissue-modality stream, each receiving 10 augmented views: 2 global whole-body views plus 8 local views, drawn preferentially from femur and lumbar-spine ROIs and padded with random small crops of the full body, so whole-body composition and fine regional bone structure share one embedding space. The objective makes every view predict the centroid of the global-view projections, with SIGReg regularization pulling the projection distribution toward an isotropic Gaussian to prevent collapse. The full pretraining bill: one L40S GPU, 400 epochs, roughly 10 hours. All downstream evaluation uses frozen embeddings with ridge or logistic probes, benchmarked against the scanner's own tabular readouts from the same scans (318 features in HPP, 117 in UKBB) and against DINOv3 ViT-H+ (about 860M parameters, internet-scale pretraining), with age, sex, and BMI as covariates throughout.
| Task | LeDXA | Baseline |
| BMD recovery (HPP, 3 clinical sites) | r = 0.882 | DINOv3 0.789 |
| Age prediction (HPP, MAE) | 3.53 yr | tabular 5.12 yr; DINOv3 4.27 yr |
| Incident hip arthrosis (UKBB, C-index) | 0.758 | tabular 0.633 |
| Incident type 2 diabetes (UKBB, C-index) | 0.753 | 0.721 |
| Osteoporosis detection (regional scans, AUROC) | 0.91 | scanner BMD 0.95; DINOv3 0.85 |
On the external cohort (47,400 UK Biobank scans, never seen in pretraining), LeDXA predicted age with MAE 2.90 years (r = 0.88). For prevalent disease, it beat the tabular baseline on 12 of 37 HPP conditions and 9 of 28 UKBB endpoints, losing on only one per cohort, and it beat DINOv3 on all eight physiological biomarkers tested (creatinine r = 0.73, hemoglobin 0.68, HDL 0.57, apnea-hypopnea index 0.54).
The longitudinal results carry the most weight. Over a median 4.3-year UKBB follow-up, LeDXA outperformed the tabular baseline on 9 of 20 incident endpoints, with the largest gap in hip arthrosis (C-index 0.758 vs 0.633). Stratifying by baseline risk, its top quartile captured 75% of year-1 hip arthrosis cases versus 43% for tabular measures, and 66% versus 41% by end of follow-up.
A biological-age gap derived from the embedding (the residual of predicted age against chronological age) tracked outcomes: the oldest-appearing quartile had 45% higher all-cause mortality hazard (HR 1.45, adjusted for age and sex), and nearly all 17 baseline conditions were more prevalent in that quartile. GWAS on the top 20 embedding components gave a mean SNP heritability of 0.143 versus 0.098 for DINOv3, recovering known body-composition and bone loci such as FTO and WNT16 plus 18 loci significant only in embedding space.
A 21.7M-parameter model trained on about 12k images in 10 GPU-hours beats an 860M generalist on this modality across disease, biomarker, mortality, and genetic alignment metrics. The lesson generalizes across medical AI: when the modality is narrow and low-dimensional, and pixel values are physical quantities, in-domain pretraining from scratch is the better deal. The applied hook is concrete. The central safety concern around GLP-1 weight-loss drugs is loss of lean mass and bone, and DXA is already the monitoring tool; a general representation that reads whole-body spatial distribution of muscle, bone, and fat gives a finer ruler for whether a weight-loss intervention is eating muscle. The biological-age gap also tracks interventions: in HPP, women's gap dropped by an average of 2.66 years after starting hormone-replacement therapy, hinting the metric is modifiable.
The authors' own list: everything is observational; both cohorts suffer healthy-volunteer selection (HPP excluded severe illness at enrollment, UKBB has the same bias); performance was not evaluated across ancestry groups; follow-up was a median 4.3 years. The medication analyses rely on self-reported use with tiny samples (11 women on HRT, 17 men on antidepressants), so they are leads, not findings. HPP disease labels are self-reported, and osteoporosis labels may not match current scans since some participants were already treated. Two further points the paper underplays: DINOv3 was used with frozen features throughout (the fine-tuning sensitivity analysis preserves the ranking, but an 860M model restricted to linear probes is somewhat handicapped), and the heritability advantage rests on a single comparison at P = 0.025.