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

In-silico trials have already proven their utility across a remarkable range of cardiovascular applications, from planning Fontan surgeries for congenital heart defects to optimizing drug delivery, stent placement, and transcatheter aortic valve implantation, widely known as TAVI. Yet the field suffers from persistent methodological weaknesses. Clinical datasets used to seed these simulations are often small, demographics are inconsistently represented, uncertainty is poorly quantified, and data transparency remains limited. A virtual trial is only as trustworthy as the population it simulates, and building a convincing virtual population requires access to hundreds or thousands of detailed anatomical models that real hospitals simply cannot provide at scale. Synthetic data generation has emerged as the most practical remedy, and machine learning methods have proven the most effective among the available techniques.

Until now, the dominant tool for statistical shape analysis in this domain has been the statistical shape model, or SSM, typically built on principal component analysis. These classical models rest on two restrictive assumptions: that shape variation is essentially linear, and that every shape can be described through point-to-point correspondence with every other shape in the dataset. Both assumptions break down when confronted with the complex, nonlinear deformations of real cardiovascular anatomy, and establishing point correspondence across patient meshes is labor-intensive and error-prone. Nonlinear extensions using kernel methods and manifold learning have partially relaxed these constraints, but they introduce heavy computational overhead and generalize poorly.

The Romanian team’s framework sidesteps these limitations entirely by adopting implicit neural representations, a family of techniques that encode three-dimensional shapes as continuous mathematical functions rather than discrete collections of points or voxels. At the heart of the approach lies a signed distance field, or SDF, a function that assigns to every point in space its distance to the object’s surface, with the sign indicating whether the point lies inside or outside. The surface itself is implicitly defined as the zero-level set of this field. Because a neural SDF is continuous and resolution-independent, it avoids the discretization biases of voxel grids, the memory overhead of volumetric representations, and the topological information loss that plagues point-cloud decoders. Unlike classical mesh-based models, it is also topology-agnostic, meaning the same architecture can represent wildly different anatomical structures.

A key technical innovation concerns how the network is trained. The seminal DeepSDF method of Park and colleagues required ground truth signed distance values, which are only well-defined for watertight surfaces with unambiguous interiors. Real aortic geometries, being open tubular structures reconstructed from cardiac CT scans, are anything but watertight. The researchers instead employed implicit geometric regularization, an approach introduced by Gropp and colleagues that exploits the fact that a true signed distance function satisfies the Eikonal partial differential equation, meaning the gradient of the field has unit magnitude everywhere. By penalizing deviations from this condition at randomly sampled points in space, and by combining it with zero-level constraints on surface points and alignment with surface normal vectors, the network learns a faithful distance field from surface samples alone, with no ground truth distance values required.

Handling non-watertight geometry demanded one further component. When the network’s zero-level set is extracted from an open surface, it produces a closed shape that includes extraneous regions beyond the anatomy of interest. To solve this, the team trained an auxiliary boundary classifier that labels each point as belonging to the inlet, the outlet, or the remaining surface. Fitted planes through the classified inlet and outlet points allow the reconstructed mesh to be precisely truncated, isolating the aortic root. The complete pipeline operates as an autodecoder: each training shape receives a trainable latent vector of 128 dimensions, and a network of roughly 1.7 million parameters, organized in eight fully connected layers with softplus activations and a skip connection, maps these latent codes and spatial coordinates to distance values. At inference, new shapes are decoded through the Marching Cubes algorithm on a 256-cubed grid, then truncated and rescaled to physical dimensions.

The framework was trained on 73 peak-systole aortic surface meshes reconstructed from cardiac CT images of patients scheduled for TAVI procedures, each clinically validated by expert cardiologists. Calcifications were represented as separate meshes, and a second, conditioned implicit network was trained to generate calcium deposits attached to a given aortic shape by receiving the aortic latent vector as an additional input. Reconstruction fidelity was measured with symmetric Chamfer distance, achieving a mean of just 0.135 square millimeters for aortic roots, evidence of high-fidelity geometric representation. New synthetic patients were produced by linearly interpolating between pairs of learned latent vectors, and from this process the team generated a validation cohort of 2,500 synthetic patients stratified by gender.

The critical question is whether these synthetic hearts pass clinical muster, and the answer appears to be yes. The researchers extracted diagnostically relevant features from both real and synthetic cohorts by slicing each mesh at half-millimeter intervals along the blood flow direction, identifying four anatomically significant cross-sections: the aortic annulus, the left ventricular outflow tract, the sino-tubular junction, and the aortic sinus. They also computed aortic valve area using a custom iterative ray-casting algorithm that converged in roughly five iterations and matched ground truth CT measurements with a mean absolute error of 0.092 square centimeters. Synthetic annulus diameters averaged 24.31 plus or minus 1.65 millimeters, closely matching the 24.6 plus or minus 2.4 millimeters reported in clinical literature for aortic stenosis populations. Statistical tests, including the Mann-Whitney U test and Welch’s t-test, found no significant differences in distribution or mean for any feature, and the distribution of aortic stenosis severity, classified as mild, moderate, or severe according to valve area thresholds, aligned well between real and synthetic cohorts.

The calcification results are perhaps the most striking. Generated calcium deposits localized to regions of mechanical stress and geometric irregularity within the conditioned aortic shape, exactly where aortic stenosis pathology develops in real patients. The deposits exhibited realistic size, morphology, and density variations that correlated with stenosis severity, and histogram analysis of 500 synthetic calcifications per shape showed spatial distributions closely matching real ground truth patterns. One acknowledged trade-off is a modest reduction in variance across most anatomical measurements, an expected consequence of latent vector interpolation when generating far more shapes than the original dataset contained. The aortic valve area, notably, preserved its variability, suggesting the primary clinical indicator remains statistically intact. The authors anticipate that clinicians will eventually generate shapes from curated samples representing specific pathological conditions rather than through random interpolation, which should further attenuate the smoothing effect.

The implications extend well beyond the aortic root. The modular architecture generalizes to other cardiovascular structures, including watertight geometries such as heart chambers, where boundary classification becomes unnecessary. The learned latent vectors can serve as compact low-dimensional representations of patient anatomy, conditioning physics-informed surrogate models that predict hemodynamic quantities like transvalvular pressure gradients, paravalvular leakage flow, and wall shear stress across entire synthetic cohorts without running full-scale fluid simulations for each geometry. This could accelerate TAVI in-silico trials, where device performance is acutely sensitive to annulus diameter, sino-tubular junction dimensions, and calcification distribution. Funded by the European Union’s Horizon 2020 program under the SIMCor project, the work points toward a future where virtual patient populations, assembled on demand and tailored to specific study designs, become a standard pillar of cardiovascular device development, reducing the cost, duration, and ethical burden of bringing life-saving technologies to patients.

Subject of Research: Generative deep learning modeling of cardiovascular geometries for in-silico clinical trials

Article Title: Generative modelling of cardiovascular geometries for in-silico trials

Article References: Gasparovici, A., & Serban, A. (2026). Generative modelling of cardiovascular geometries for in-silico trials. Neural Computing and Applications, 38(18), Article 745. https://doi.org/10.1007/s00521-026-12463-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12463-8

Keywords: implicit neural representations, signed distance field, in-silico trials, cardiovascular modeling, aortic root, synthetic data generation, TAVI, aortic stenosis, statistical shape models, calcifications, deep learning, virtual patients