Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability

LeDXA: Self-Supervised Vision Model Unlocks Systemic Health Insights from Routine DXA Scans

Dual-energy X-ray absorptiometry (DXA) is a cornerstone of clinical screening. It is routinely used to measure bone mineral density (BMD) and regional body composition. However, the vast majority of the spatial information within these scans remains untapped. Most clinicians discard this data in favor of a few simplified numerical readouts. Researchers have now developed an AI model called LeDXA. This model looks at standard scans to predict much more than just bone health. By analyzing the spatial structure of the images, the model can estimate biological age. It can also predict risks for diseases like diabetes and arthritis. It can even identify genetic patterns.

The Information Gap in Clinical Imaging

Current clinical practice treats DXA scans as a source of scalar summaries. A technician runs a scan. The software then outputs a handful of numbers. These include a regional bone density average or a total fat mass. They might also include visceral adipose tissue (VAT) volume (the fat stored around internal organs). While these metrics are useful, they represent a massive compression of the actual data. Standard measurements reduce a complex 2D spatial map into a single digit. This process strips away the nuances of bone texture. It also ignores the precise geometry of fat deposition. It misses the subtle distribution of lean tissue across the limbs.

This compression creates a blind spot. Certain indicators like bone density are excellent for diagnosing osteoporosis. However, they fail to capture the systemic, body-wide variations that drive chronic diseases. Previous attempts to bridge this gap have relied on supervised learning. This involves training models to recognize specific labels like "fracture risk." This approach is limited. It requires expensive, manually labeled clinical data for every new disease of interest. Furthermore, general-purpose vision models like DINOv3 are suboptimal for this task. They are tuned for the edges and colors of natural photographs. They are not designed for the physical density distributions inherent to X-rays. Consequently, they struggle to extract deep biological signals from medical imagery.

Predicting Latents via the LeJEPA Framework

To move beyond pixel-level reconstruction, the researchers introduced LeDXA. This is a vision foundation model built on a joint-embedding predictive architecture (JEPA). Traditional generative models try to reconstruct every pixel of an image. This task often wastes computational energy on irrelevant scanner noise. Instead, LeJEPA learns by predicting representations in a latent space (a mathematical space representing high-level features). Rather than redrawing the image, the model predicts the "meaning" of different parts of the scan.

The architecture follows a structured self-supervised pipeline :

1. Multi-view Augmentation: Each raw DXA scan is expanded into ten different views. These include two "global" views of the entire body. They also include eight "local" views consisting of smaller crops. These crops include detailed shots of the femur (thigh bone) or lumbar spine.
2. Shared Encoding: A Vision Transformer (ViT-Small/16) acts as the encoder. It processes these various views. It transforms them into a 384-dimensional mathematical representation known as an embedding.
3. Latent Prediction: The model is trained to push the representations of the local crops toward the average of the global views. This forces the model to understand how local anatomical structures relate to the overall body composition.
4. Regularization: The authors used Sketched Isotropic Gaussian Regularization (SIGReg) to prevent "representational collapse." Collapse is a failure mode where the model outputs the same constant value for every image. SIGReg ensures the distribution of the learned embeddings remains mathematically stable.

The model was trained from scratch on 11,540 unlabeled scans from the Human Phenotype Project. It learns to navigate human anatomy without needing human labels.

Evidence of Systemic Signal Capture

LeDXA outperforms both traditional tabular measurements and general models. It does so despite having significantly fewer parameters and training images. The authors report that LeDXA achieves a mean absolute error (MAE) for chronological age prediction of 3.53 years. This is a major improvement over the 5.12 years provided by standard tabular DXA measures [Figure 2a].

The model's true strength lies in its ability to act as a multi-system health sensor. In a cross-sectional analysis, LeDXA predicted a wide array of physiological biomarkers. These included creatinine (a kidney function marker), hemoglobin (an indicator of blood health), and HDL cholesterol [Figure 2b]. When classifying prevalent diseases, the model showed significant improvements over the tabular baseline. These gains spanned multiple organ systems, including cardiovascular, metabolic, and respiratory domains [Figure 2c].

The model's longitudinal performance is perhaps most striking. Using data from the UK Biobank, the authors measured how well LeDXA predicts the incidence (new occurrences) of future diseases. This was done over a median 4.3-year follow-up. The model demonstrated superior discrimination for joint-related conditions. For instance, in predicting hip arthrosis, LeDXA achieved a concordance index (a measure of predictive accuracy) of 0.758. This is much higher than the 0.633 achieved by traditional tabular measures [Figure 3a]. This means 66% of incident hip arthrosis cases occurred in the highest-risk quartile identified by LeDXA. In contrast, tabular measures only caught 41% of these cases [Figure 3b].

Limitations and Observational Boundaries

Despite these impressive capabilities, the study has constraints. First, all findings are purely observational. The model identifies correlations between morphological patterns and disease states. It cannot establish causality. It cannot prove that a specific fat distribution causes diabetes. It only shows that they appear together in the image.

Second, the analysis of medication and disease status relies on self-reported data. This introduces potential bias. For example, the analysis of the "biological age gap" follows hormone-replacement therapy (HRT). This analysis relies on participants reporting their own medication use. Such reports may be subject to recall errors. Finally, the datasets used are prone to "healthy-volunteer" selection bias. Participants in the Human Phenotype Project and the UK Biobank tend to be more health-conscious. Therefore, the model's ability to generalize to highly diverse clinical populations remains to be tested.

The Verdict: A New Lens for Precision Medicine

The evidence suggests that LeDXA is a successful proof-of-concept. It proves that specialized self-supervised architectures can extract profound biological truths. This is possible even with relatively small datasets. The model outperforms much larger general-purpose models. By uncovering a "biological age gap," LeDXA connects morphology to mortality risk .

It also tracks responses to medications. This moves DXA from a static measurement of density to a dynamic tool for monitoring health.

Integrating these high-dimensional embeddings into clinical workflows is the next challenge. Physicians must be able to use this data without increasing their workload. For researchers, the door is now open. We can use DXA images to map the complex, heritable, and interconnected landscape of human health. Code and implementation details are available at https://github.com/GilSasson1/LeDXA.

Figures from the paper