New Dual-Level Test Reveals When Fake Anatomy Looks Real but Isn’t

Synthetic anatomies are quietly becoming one of the most powerful tools in modern medicine. When surgeons rehearse a complex spinal procedure on a virtual patient, when an artificial intelligence system learns to segment organs from scarce CT scans, or when researchers run entire clinical trials inside a computer, they rely on artificially generated three-dimensional anatomical models that must look, and crucially behave, like real human bodies. A new study published in Machine Learning with Applications argues that the way scientists judge whether these synthetic bodies are trustworthy has been fundamentally incomplete, and it proposes a two-tier testing system that exposes hidden flaws in the most popular generation methods.

The research, led by Luca Di Angelo, Emanuele Guardiani, Tamsir Jobe, Ivan Letteri, Antonio Marzola and Pierpaolo Vittorini, addresses a deceptively simple question: when a computer generates a fake vertebra, is it actually anatomically plausible, or does it merely look plausible on the metrics we happen to be measuring? The team’s answer is a dual-level evaluation framework that separates two properties which most previous studies have conflated. The first level measures geometric-statistical fidelity, asking how closely the spatial distribution of synthetic shapes matches the real training data. The second level measures morphometric plausibility, asking whether the synthetic samples preserve the multivariate relationships among clinically meaningful measurements, such as vertebral body heights, canal widths, endplate areas and facet joint geometry, that define real human anatomy.

The need for such a framework arises from a peculiar blind spot in the field. Data augmentation, the practice of expanding small medical datasets with synthetic samples, has become standard practice precisely because expert annotation of three-dimensional anatomy is slow, expensive and constrained by privacy regulation. Yet the overwhelming majority of studies judge augmented data solely by whether it improves downstream performance, such as segmentation or classification accuracy. The authors point out that this is a poor proxy for anatomical realism. A generative model could produce samples riddled with anatomically impossible proportions while still boosting a classifier’s score, because the model exploits dataset-specific patterns rather than genuine biological variability. Conversely, a method producing beautifully realistic anatomy might yield no measurable performance gain at all.

To test their framework, the researchers chose the second lumbar vertebra, the L2, as a case study, drawing on the publicly available VerSe database of annotated CT scans. They extracted 97 three-dimensional vertebral surface models, partitioning them into a development set of 32 shapes for augmentation and training, an independent reference set of 60 real vertebrae, and a small external test set of 5. Three representative augmentation strategies were compared: SMOTE-style interpolation between neighboring shapes in geometric space, a variational autoencoder that learns a probabilistic latent representation of vertebral form, and a statistical shape model that samples principal modes of population-level shape variation. Each method generated 100 synthetic vertebrae.

The first level of evaluation used a complementary pair of metrics. Maximum Mean Discrepancy, computed in a reproducing kernel Hilbert space, quantified how far the overall distribution of synthetic vertex configurations strayed from the real distribution, while the Chamfer distance measured local geometric proximity between point clouds. The results were already revealing. The statistical shape model achieved the best agreement with the global distribution of the training data but a comparatively higher Chamfer distance. The variational autoencoder did the opposite, achieving the closest local geometric fit while deviating most from the global statistics. SMOTE interpolation landed in between, balancing both dimensions. On their own, these numbers would suggest the interpolation method was the most faithful generator.

Then came the second level, and the story changed dramatically. The team extracted 16 morphometric descriptors per vertebra, covering lengths, widths, heights, volumes, areas and angles measured within a rigorously defined anatomical coordinate system, after using Pearson correlation analysis to collapse redundant features into anatomically meaningful composites. Each synthetic vertebra was then assigned a squared Mahalanobis distance from the reference distribution of the 60 real vertebrae, a multivariate measure that accounts for the correlations among descriptors, so that an implausible combination of individually normal measurements cannot slip through. Plausibility thresholds were derived empirically from the real population using a leave-one-out procedure, classifying each sample as typical, borderline or an outlier.

The outcome exposed a striking dissociation. Every synthetic vertebra from every method remained anatomically plausible, with no outliers, but the coverage of anatomical varied enormously. SMOTE interpolation and the variational autoencoder both produced samples concentrated in the central region of the morphometric space, underrepresenting the peripheral anatomical configurations seen in the broader real population. The statistical shape model, by contrast, spread its samples across a much wider morphometric range that more closely matched the variability envelope of the real cohort while staying within plausible bounds. In other words, the method that looked best on geometric similarity was not the one that best reproduced the rich diversity of actual human anatomy.

Downstream experiments reinforced the message. Training a random forest classifier to label the nine anatomical regions of each vertebral surface mesh, the researchers found that the statistical shape model delivered the strongest and most consistent test accuracy, climbing steadily from 0.969 to 0.975 as synthetic sample size grew from 25 to 100. SMOTE produced modest, non-monotonic gains that peaked at 75 synthetic samples and slightly degraded at 100, a signature of overfitting to a concentrated region of shape space. The variational autoencoder barely improved on the baseline at any sample size, consistent with its habit of generating tightly clustered, low-variability samples. Crucially, the method ranking on downstream performance did not match the ranking on geometric fidelity, confirming that neither criterion alone tells the full story.

The implications extend well beyond the spine. The authors deliberately designed the framework to be domain-independent: it applies to any anatomical structure for which a consistent shape representation and validated morphometric descriptors exist, from blood vessels to cardiac tissue. This matters enormously for in silico trials, the emerging paradigm in which virtual patient cohorts replace or supplement real clinical studies to test medical devices and interventions. A virtual cohort generated by a method that concentrates samples near the center of the anatomical distribution would systematically underrepresent extreme morphologies, precisely the patients most likely to experience device failure, introducing a subtle and dangerous selection bias that no amount of geometric checking would reveal.

The study also acknowledges its limits. The framework requires a stable, well-conditioned morphometric reference space, and in very small data regimes the covariance estimation underpinning the Mahalanobis distance may need additional regularization or dimensionality reduction. The authors plan to extend the framework to further anatomical structures and to newer hybrid generative approaches that combine statistical shape models with deep learning. But the central lesson is already clear and likely to reshape how medical AI practitioners evaluate their synthetic data: realism is not one property but two, and a fake vertebra that perfectly mimics the geometry of its training set can still be, in the sense that matters most for patients, anatomically wrong. Choosing an augmentation strategy, the authors argue, should depend on the goal, whether reinforcing local training distributions or generating genuinely diverse virtual populations, and only a dual-level evaluation can tell them apart.

Subject of Research: Evaluation of anatomical data augmentation methods for 3D vertebra segmentation

Article Title: A Dual-Level Evaluation Framework for Anatomical Data Augmentation: A Comparative Study on L2 Vertebra Semantic Segmentation

Article References: Angelo, L. D., Guardiani, E., Jobe, T., Letteri, I., Marzola, A., & Vittorini, P. (2026). A Dual-Level Evaluation Framework for Anatomical Data Augmentation: A Comparative Study on L2 Vertebra Semantic Segmentation. Machine Learning with Applications, Article 101019. https://doi.org/10.1016/j.mlwa.2026.101019

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101019

Keywords: data augmentation, medical imaging, anatomical modeling, statistical shape model, variational autoencoder, SMOTE, L2 vertebra, semantic segmentation, Mahalanobis distance, morphometric plausibility, in silico trials, machine learning