PRISM: A 3D Probabilistic Neural Representation for Interpretable Shape Modeling

AI summary

Overview

• Research area: Statistical shape modeling, neural implicit representations, uncertainty quantification, and computational anatomy for medical imaging (with an application to pediatric airway anatomy).
• Technical level: Advanced. The paper combines neural implicit fields with information geometry (Fisher Information, Cramér-Rao bounds), Cholesky covariance parameterization, and diffeomorphic deformation concepts.
• Scope (one sentence): PRISM models the conditional distribution of anatomical shape given a covariate time as a spatially heteroscedastic Gaussian field over a shared template, and derives a closed-form Fisher Information metric that yields analytic, pointwise temporal uncertainty at arbitrary spatial resolution.

What This Paper Is About

Existing covariate-aware shape models are either deterministic (they give a single predicted deformation without any confidence estimate) or they represent variability in a deformation parameter space (such as LDDMM momenta or time shifts), from which a pointwise uncertainty map on the anatomy cannot be obtained analytically. PRISM addresses this gap by treating shape change as a continuous Gaussian field defined directly on the template and by deriving a closed-form Fisher Information expression that converts the learned mean trajectory and covariance into a spatially varying estimate of temporal (developmental) uncertainty. The goal is a unified framework that supports shape trajectory modeling, individual developmental-stage inference, personalized prediction, and anomaly detection, while producing interpretable uncertainty estimates that are clinically meaningful.

Key Contributions

1. A conditional probabilistic implicit field that jointly models the mean developmental trajectory and spatially varying population variability from cross-sectional (not necessarily longitudinal) data.
2. A closed-form Fisher Information metric that enables efficient, analytically tractable local temporal uncertainty quantification through automatic differentiation, avoiding Monte Carlo propagation and its sampling variance.
3. An amortized inverse encoder that estimates intrinsic developmental time from local shapes without test-time optimization, supporting dense per-point temporal inference across the anatomy.
4. A unified set of applications — population shape trajectories, intrinsic time inference, personalized longitudinal shape prediction, and out-of-distribution (OOD) detection — evaluated on three synthetic datasets and one clinical dataset.

Main Findings

• Mean trajectory reconstruction is competitive or superior. On the Airway dataset, PRISM attains the lowest reconstruction error of the compared methods (CD 0.064, HD 9.614, EMD 1.308, metrics scaled by 100, lower is better), compared with NAISR (0.072, 10.075, 1.422) and A-SDF (0.114, 10.508, 2.040). On Starman (G), Starman (L), and ANNY it remains comparable to the best performer. PRISM improves on nearly all metrics against NAISR across all datasets; the authors attribute this to PRISM's use of precomputed dense correspondences, which decouples correspondence from reconstruction.
• A-SDF overfits on the smaller clinical dataset. Although A-SDF fits the mean trajectory well with sufficient data (Starman G&L, ANNY), it yields higher reconstruction errors than the deformation-based methods on the Airway dataset.
• Amortized intrinsic time estimation is far faster and highly accurate on synthetic data. On Starman (G), PRISM (Amortized) achieves r = 1.000, R² = 0.999, MAE = 0.005, and 0.040 s per case, versus A-SDF (r = 0.992, R² = 0.983, MAE = 0.016, 4.005 s) and NAISR (r = 0.991, R² = 0.980, MAE = 0.019, 7.892 s), both of which use test-time optimization (TTO). The paper describes amortized PRISM as remaining competitive with baselines while being an order of magnitude faster.
• On Airway, a TTO variant of PRISM is best. With test-time optimization, PRISM reaches r = 0.923, R² = 0.843, and MAE = 19.618 months, outperforming NAISR TTO (r = 0.908, R² = 0.820, MAE = 20.512 months) and A-SDF TTO (r = 0.745, R² = 0.334, MAE = 42.886 months). The amortized PRISM is slightly behind NAISR TTO on this dataset (r = 0.893, R² = 0.792, MAE = 22.639 months, 0.805 s per case), which the paper attributes to the amortization trade-off rather than a methodological disadvantage.
• PRISM is the only method applicable to spatially varying intrinsic time. Baselines estimate a single global time per shape, so on Starman (L), where arms and legs follow distinct developmental trajectories, PRISM achieves near-perfect estimation (arm: r = 1.000, R² = 0.999, MAE = 0.008; leg: r = 1.000, R² = 0.999, MAE = 0.004).
• Uncertainty estimates align with ground truth on synthetic data. Figures 2 and 3 report tight alignment between PRISM's estimated conditional distribution of intrinsic time and the ground-truth distribution on Starman (G) and Starman (L), and show that decoding shapes at the mean and ±2σ of the predicted intrinsic time distribution produces contours that overlap closely with ground-truth shapes.
• ANNY results. PRISM amortized achieves r = 0.993, R² = 0.971, and MAE = 0.602 years at 0.430 s per case, compared with A-SDF (r = 0.996, R² = 0.991, MAE = 0.351 years, 7.045 s) and NAISR (r = 0.988, R² = 0.958, MAE = 0.933 years, 8.606 s).
• Not reported in the provided content. The text supplied ends partway through the temporal uncertainty validation on Starman (G&L); results for the clinical uncertainty figure (Figure 4) and the personalized prediction and OOD detection experiments are not included in the available content.

Methodology in Plain English

PRISM works on shapes represented as displacement fields from a shared template: every subject's anatomy is described by how far each template point moves to reach that anatomy, which requires point correspondence across subjects (established by a template-based registration module the authors describe as outside their primary contribution and defer to Appendix A.1).

For each template point p and covariate time t, PRISM models the displacement d as a Gaussian with a mean μ(p,t) and a covariance Σ(p,t). Because the template point is fixed, modeling the displacement is equivalent to modeling the target position up to a deterministic translation, and the covariance is unchanged. Both μ and the Cholesky factor L of Σ (with Σ = LLᵀ) are produced by coordinate-based neural networks. Each is written as a covariate-driven term plus a covariate-independent residual — [f(p,t) − f(p,0)] + h(p) — so that the covariate-driven part vanishes at t = 0, which enforces identifiability between the two components. For 3D shapes, the Cholesky factor has 6 free parameters per query point (3 for 2D geometry); a softplus activation on diagonal entries plus a small positive constant keeps Σ symmetric positive definite.

Training deliberately avoids the known bias from jointly optimizing mean and variance under a Gaussian negative log-likelihood. The mean branch is trained with an ℓ1 loss (for robustness to outliers) and the covariance branch with the negative log-likelihood while the mean is held fixed. A two-stage curriculum first freezes the covariance head and trains only the mean for T_warm = 10 epochs, then trains both branches jointly.

To estimate an individual's intrinsic developmental time (distinct from chronological time, since subjects of the same age may be developmentally advanced or delayed), the paper poses a maximum likelihood problem but solves it with amortized inference: an inverse encoder is trained with an L1 loss on synthetic triplets generated by sampling template coordinates and intrinsic times uniformly and querying the learned forward model for the corresponding mean displacement. This avoids iterative per-subject optimization.

Uncertainty comes from information geometry. The Fisher Information of the Gaussian field has a closed form with two terms: one from how the mean changes with time, and one from how the covariance changes with time. The authors keep only the mean term, because the two are orthogonal under the Fisher-Rao metric and the covariance term answers a different question (how anatomical diversity evolves, rather than how precisely an individual can be localized along the trajectory). Under the assumption that the population-average intrinsic time equals chronological time, the Cramér-Rao inequality bounds the variance of intrinsic time below by the inverse Fisher Information, and PRISM uses that inverse as its temporal uncertainty estimate. Because the mean's time derivative comes from automatic differentiation and the inverse covariance from a single forward pass, the estimate is analytic and can be queried densely across the anatomy without sampling.

The Fisher Information is also used downstream: it weights each point's intrinsic time estimate when aggregating to a global progression estimate (so that temporally discriminative regions count more), it defines a temporal z-score used to hold a subject's developmental stage constant when forecasting future shape, and it normalizes a per-point z-score-like OOD score that flags regions appearing developmentally younger than the rest of the same anatomy — aimed at pathology such as subglottic stenosis without labeled anomaly data.

Why This Matters

• Impact on research: PRISM offers a route to uncertainty quantification in covariate-conditioned shape modeling that is analytic rather than sampling-based, producing a pointwise field on the anatomy rather than statistics confined to a deformation parameter space. It also shows a way to learn from cross-sectional clinical data where complete closed surfaces and longitudinal sequences are unavailable.
• Clinical growth and development tracking: producing spatially resolved normative ranges over age, distinguishing developmentally conserved anatomy from naturally diverse anatomy.
• Personalized longitudinal prediction: forecasting a child's future anatomy while preserving whether they are developmentally advanced or delayed relative to the population mean.
• Anomaly detection without labeled anomalies: flagging airway regions that lag behind the rest of the anatomy, applicable to conditions such as subglottic stenosis, using only the likelihood of an observed shape under the learned population distribution.
• Industry relevance: the amortized inverse encoder reduces per-case inference to fractions of a second (for example, 0.040 s on Starman G versus 7.892 s for NAISR with test-time optimization), making near-real-time shape analysis and developmental staging plausible in clinical pipelines; the released code at https://github.com/uncbiag/PRISM supports reproducibility and downstream deployment.

Future Directions

• Reporting the remaining experiments. The provided content stops before the personalized prediction and OOD detection results; evaluating whether the OOD score separates the 31 subglottic stenosis scans from the 358 normal-airway CT scans is the natural next question.
• Reducing the amortization gap. On the clinical Airway dataset, amortized PRISM trails NAISR with test-time optimization (r = 0.893 versus 0.908; MAE 22.639 versus 20.512 months), while PRISM with test-time optimization leads (r = 0.923). Closing the gap between amortized and test-time-optimized inference is an open direction.
• Extending beyond intact template correspondence. The method depends on point correspondence established by a separate registration module; how robust PRISM is to imperfect correspondence or incomplete observations is not established in the reported experiments.
• Removing the population-mean-time assumption. The Fisher Information formulation assumes that the population-average intrinsic time equals chronological time, and the authors retain only the mean-trajectory term of the Fisher Information, discarding the covariance term that captures how structural variability changes over time — a term whose independent value is left unexplored.

Target Audience

Readers who will benefit most are machine learning and medical imaging researchers working on statistical shape modeling, neural implicit representations, and uncertainty quantification; computational anatomy groups studying growth and development; and clinical researchers in pediatric otolaryngology or airway analysis who need spatially localized normative references and anomaly detection tools. A working familiarity with Gaussian models, Cholesky factorization, and information geometry is helpful, as the paper is written at an advanced technical level.

Authors’ abstract

Understanding how anatomical shapes evolve in response to developmental covariates - and quantifying their spatially varying uncertainties - is critical in healthcare research. Existing approaches typically rely on global time-warping formulations that ignore spatially heterogeneous dynamics. We introduce PRISM, a novel framework that bridges implicit neural representations with uncertainty-aware statistical shape analysis. PRISM models the conditional distribution of shapes given covariates, providing spatially continuous estimates of both the population mean and covariate-dependent uncertainty at arbitrary locations. A key theoretical contribution is a closed-form Fisher Information metric that enables efficient, analytically tractable local temporal uncertainty quantification via automatic differentiation. Experiments on three synthetic datasets and one clinical dataset demonstrate PRISM's strong performance across diverse tasks - from modeling shape evolution to personalized shape prediction and anomaly detection - within a unified framework, while providing interpretable and clinically meaningful uncertainty estimates.