Med Tech , AI in Pharma , AI in Life sciences Aug 07, 2026

Shape Forecasting as a Clinical Asset: What Brain Morphometry AI Means for Trial Design and Prognosis Infrastructure

VECTOR Labs Team

Last updated on: Aug 07, 2026

The dominant narrative in healthcare AI still centres on classification: does this scan show disease or not? That framing has produced useful tools, but it misses a more commercially significant problem for firms building clinical platforms. The question that actually moves trial economics and prognosis infrastructure is not what a brain structure looks like today, but how it will deform over the next one to three years. Continuous-time mesh prediction models for subcortical structures are now reaching a level of technical maturity that makes this question answerable in production systems, and the engineering decisions required to support them are different enough from diagnostic AI that CTOs need to get ahead of them now.

Why Shape Trajectories Carry More Signal Than Volume Metrics

Volume-based atrophy measures have been the standard quantitative endpoint in neurodegenerative trials for years. They are tractable, reproducible, and well understood by regulators. The problem is that volume collapses a three-dimensional deformation pattern into a single scalar, discarding the spatial distribution of tissue loss that distinguishes, for example, hippocampal atrophy patterns across Alzheimer's subtypes.

Surface-based morphometry preserves that spatial information by tracking the position and geometry of thousands of vertices across a structure's mesh over time. This means a model can learn that atrophy is progressing faster at the CA1 subfield than at the subiculum, and use that pattern to predict future shape rather than simply extrapolating a volume trend.

The clinical implication is that shape trajectory models can stratify patients who look identical on volumetric measures but who will diverge substantially in cognitive outcomes over a two-year trial window. That stratification capability is what makes these models interesting for trial enrichment, not just academic morphometry research.

The Architecture: What MT-GNN Gets Right

Metric Tensor Prediction

Most longitudinal mesh predictors take one of two approaches: they either regress trajectories through a shape embedding space and extrapolate forward, or they learn to deform vertices directly from scan-to-scan differences. Both approaches struggle against a deceptively strong baseline, which is simply the temporal mean of observed visits. The temporal mean produces a plausible denoised shape, and because true per-visit shape change is small relative to pipeline noise, a model that cannot beat it is not learning longitudinal signal at all.

MT-GNN takes a different route by predicting the surface's intrinsic geometry rather than its extrinsic vertex positions. Specifically, it predicts the per-vertex first fundamental form (the metric tensor) for an arbitrary visit history and an arbitrary prediction horizon, conditioned on a Fourier encoding of the lead time. The predicted metric is then decoded into a surface via a differentiable As-Rigid-As-Possible solver (Ding et al., arXiv 2026).

End-to-End Training Through Reconstruction

The solver is differentiable, which means the entire pipeline from metric prediction to surface reconstruction is trained end-to-end on vertex error. This matters because training through reconstruction constrains the predicted geometry to produce valid surfaces, not just numerically plausible metric tensors. On 14 subcortical structures from the ADNI dataset, MT-GNN achieves a mean vertex error 2.29% below the temporal mean baseline, beating it on all 14 structures, compared to 0.19% for geodesic shape regression and 0.45% for a mesh transformer, with the performance gap widening at longer prediction horizons (Ding et al., arXiv 2026).

The widening gap at longer horizons is the practically important result. Trial endpoints are typically assessed at 18 to 24 months. A model that degrades relative to baseline at that range is not useful for enrichment. MT-GNN's continuous-time conditioning via Fourier lead-time encoding is what allows it to maintain accuracy across arbitrary horizons without retraining for each timepoint.

Clinical Trial Enrichment: What This Changes

Cohort Selection

Trial enrichment using shape trajectory models works by screening candidates not just on current diagnosis or current volume, but on predicted rate of structural change over the trial period. Subjects predicted to show rapid, spatially specific atrophy are more likely to show detectable treatment effects against an active comparator. This reduces the sample size required to power a trial at a given effect size, which is a meaningful cost lever when Phase III neurodegeneration trials routinely run to thousands of participants.

The practical requirement is that enrichment predictions must be generated pre-randomisation, from baseline and any available prior scans. This means the model must handle variable visit histories gracefully, including single-visit baselines where no prior trajectory exists. MT-GNN's causal multi-visit conditioning architecture is designed for exactly this regime, but single-visit cold-start performance is something any production deployment team should validate explicitly on their own cohort before relying on it for enrichment decisions.

Endpoint Design

Shape trajectory predictions can also inform endpoint selection. If a model predicts that a specific region of the hippocampal surface will show the greatest deformation in a given patient subgroup, that localised change can be pre-specified as a primary or secondary endpoint rather than whole-structure volume. Pre-specification based on model predictions requires careful statistical planning to avoid inflating false positive rates, but it is a legitimate design strategy when the prediction is generated from held-out data and the endpoint is locked before unblinding.

Longitudinal Data Requirements and the Infrastructure Gap

What the Training Pipeline Needs

These models require longitudinal MRI data with consistent acquisition protocols, segmentation pipelines, and surface reconstruction methods across visits. ADNI provides this by design. Most health-tech platforms do not. Retrospective data collected across scanners, field strengths, and software versions introduces systematic mesh distortions that a model trained on harmonised data will interpret as longitudinal signal rather than acquisition noise.

Before committing to a shape trajectory modelling capability, engineering teams need to audit their longitudinal data for protocol consistency, assess whether harmonisation methods such as ComBat or deep learning-based scanner normalisation are sufficient to reduce cross-site variance below the per-visit shape change signal, and establish a minimum acceptable visit interval and visit count for training and inference. These are not model architecture questions. They are data infrastructure questions that determine whether the architecture ever has a chance to work.

Inference and Serving Constraints

Graph neural networks operating on high-resolution subcortical meshes are not lightweight inference workloads. A single structure prediction at hippocampal mesh resolution involves thousands of vertices and their neighbourhood relationships. Serving predictions for a trial cohort of several hundred subjects across 14 structures at multiple prediction horizons requires careful batching strategy and hardware allocation. The differentiable ARAP solver adds reconstruction overhead on top of the GNN forward pass. Teams should benchmark end-to-end latency and throughput requirements against their trial monitoring cadence before finalising serving architecture.

Regulatory and Validation Considerations

Shape trajectory prediction sits in a different regulatory category than diagnostic classification. A model that predicts future atrophy rate and influences trial enrolment is functioning as a prognostic device, and depending on jurisdiction and how the output is used clinically, it may require SaMD classification and performance validation against a pre-specified reference standard. The absence of a ground truth future scan at the time of enrichment prediction means validation must rely on prospective follow-up data, which extends the timeline for regulatory submission substantially.

Validation datasets should be drawn from populations that match the intended trial indication, not just from ADNI, which over-represents well-characterised Alzheimer's disease. Models validated only on ADNI cohorts may not generalise to early-stage disease, atypical presentations, or non-Western populations. Subgroup performance analysis across age, sex, APOE status, and disease stage should be treated as a minimum validation requirement, not an optional analysis.

The broader principle is that the model's commercial value depends entirely on whether its predictions are trustworthy enough to influence enrolment decisions. That trustworthiness is established through validation rigour, not through benchmark performance on a held-out ADNI split.

Where Vector Labs Fits

We build and certify production AI systems for clinical applications, including model validation pipelines structured to meet medical device software standards. Our work developing and certifying a cardiac AI model for atrial fibrillation detection from wearable ECG signals, resulted in Class 2A medical device certification and demonstrates the validation and documentation approach we apply to prognostic model development. If you are scoping a shape trajectory modelling capability or a trial enrichment platform, contact us at vector-labs.ai/contacts.

FAQs

How many longitudinal scans does a model like MT-GNN require per subject to produce reliable predictions?

MT-GNN is designed to condition on an arbitrary causal visit history, which means it can technically generate predictions from a single baseline scan. However, single-visit predictions carry substantially higher uncertainty because there is no observed trajectory to anchor the extrapolation. In practice, two to three prior scans spaced at least six months apart provide enough longitudinal signal to produce predictions that are meaningfully more informative than the temporal mean baseline. Teams should validate single-visit cold-start performance explicitly on their own cohort before using it for enrichment decisions.

What does it mean in practice to predict the metric tensor rather than vertex positions directly?

The metric tensor encodes the intrinsic geometry of a surface at each vertex: specifically, how distances and angles are locally distorted relative to a reference configuration. Predicting the metric tensor rather than absolute vertex positions means the model is learning how the surface is stretching and compressing over time, not just where each point ends up. This is a more geometrically stable target because it is invariant to the rigid alignment applied during preprocessing, and because the differentiable ARAP solver reconstructs a valid surface from the predicted geometry rather than allowing the decoder to produce anatomically implausible shapes.

How should we think about data harmonisation requirements before training on multi-site MRI data?

The core problem is that scanner-induced differences in image contrast, resolution, and geometric distortion propagate into segmentation and surface reconstruction outputs, creating apparent shape differences that have nothing to do with biology. If these differences are larger than the longitudinal shape change signal the model is trying to learn, the model will fit noise rather than trajectory. Harmonisation methods such as ComBat can reduce scanner effects on derived morphometric features, but they operate on summary statistics rather than raw meshes. Deep learning-based image harmonisation applied before segmentation is a more principled approach for mesh-level prediction, though it introduces its own validation burden. The minimum requirement is to quantify cross-site mesh variance on test-retest or travelling phantom data and confirm it is below the expected per-year atrophy magnitude for your target population.

What regulatory pathway applies to a shape trajectory model used for clinical trial enrichment?

The regulatory classification depends on how the model output is used. If predictions are used to inform enrolment decisions that affect individual patient access to an investigational treatment, most regulatory frameworks will treat the software as a Software as a Medical Device requiring formal SaMD classification and performance validation. In the US, this falls under FDA's SaMD guidance and potentially the De Novo or 510(k) pathway depending on the risk classification. In the EU, MDR 2017/745 applies. The key validation challenge is that the reference standard (the future scan) is not available at the time of prediction, so prospective follow-up data is required to establish clinical validity, which adds 12 to 24 months to a typical regulatory timeline.

Can shape trajectory models be applied to structures beyond the hippocampus, and does performance vary significantly across structures?

MT-GNN was evaluated across 14 subcortical structures from the ADNI dataset and outperformed competing methods on all 14, which suggests the architecture generalises across structures rather than being tuned to hippocampal geometry specifically. That said, performance will vary with the magnitude of longitudinal shape change relative to segmentation noise for each structure. Structures with slower or more spatially diffuse atrophy patterns will produce smaller performance margins over the temporal mean baseline. Teams targeting specific structures for a given indication should evaluate per-structure performance on a held-out cohort from their own data distribution rather than relying on ADNI benchmarks alone.

What are the main infrastructure differences between deploying a shape trajectory model versus a standard diagnostic classification model?

Diagnostic classification models typically operate on a single scan at inference time and return a scalar or categorical output. Shape trajectory models require a longitudinal record of prior scans, consistent preprocessing applied across all visits, and a serving architecture that can handle graph-structured inputs at mesh resolution. The ARAP reconstruction step adds compute overhead beyond the GNN forward pass. Storing and versioning per-subject mesh histories also creates data management requirements that are more complex than storing single-scan feature vectors. Teams should also plan for model retraining or fine-tuning as their longitudinal dataset grows, since trajectory models benefit more from additional follow-up data than classification models do from additional cross-sectional cases.

A team that understands you

With 20+ years of experience in the world's leading consultancy companies, implementing AI and ML projects in industry-specific contexts, we are ready to hear your challenges.

Talk with an AI expert