🤖 machine learning

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

The paper introduces PRISM, a novel framework that combines implicit neural representations with uncertainty-aware statistical analysis to model the conditional distribution of anatomical shapes given covariates, enabling spatially continuous, interpretable uncertainty quantification and outperforming existing global time-warping methods across diverse healthcare tasks.

Original authors: Yining Jiao, Sreekalyani Bhamidi, Carlton Jude Zdanski, Julia S Kimbell, Andrew Prince, Cameron P Worden, Samuel Kirse, Christopher Rutter, Benjamin H Shields, Jisan Mahmud, Marc Niethammer

Published 2026-06-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yining Jiao, Sreekalyani Bhamidi, Carlton Jude Zdanski, Julia S Kimbell, Andrew Prince, Cameron P Worden, Samuel Kirse, Christopher Rutter, Benjamin H Shields, Jisan Mahmud, Marc Niethammer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a child's body grows from infancy to adulthood. You have photos of hundreds of different kids at different ages. You want to answer two big questions:

What is the "average" way a body part (like an arm or an airway) changes shape as time passes?
How much does that shape vary from person to person at any given moment? Is it a tight, predictable change, or is it a wild, unpredictable swing?

Most existing computer models try to answer the first question but ignore the second. They say, "At age 5, the average arm looks like this," but they can't tell you how much the arm might wiggle or vary for different kids. Other models try to guess the variation, but they do it so slowly and clumsily that they can't give you a clear answer for every single spot on the body.

PRISM is a new computer framework that solves this. Think of it as a "3D Crystal Ball for Human Growth" that doesn't just predict the future; it also tells you how confident it is in that prediction.

Here is how it works, using some simple analogies:

1. The "Stretchy Map" (Implicit Neural Representation)

Imagine you have a standard, perfect map of a human body (a template). Now, imagine that every real person is just this map stretched, squished, or twisted in specific ways.

Old models tried to guess the stretch for the whole body at once.
PRISM treats the body like a stretchy, elastic fabric. It learns exactly how to stretch that fabric for a 5-year-old, a 10-year-old, or a 15-year-old. But instead of just giving you one "average" stretch, it gives you a cloud of possibilities. It says, "At this spot on the elbow, the stretch could be this much, or that much, and here is the probability of each."

2. The "Confidence Meter" (Fisher Information)

This is the paper's biggest trick. Usually, to figure out how uncertain a model is, you have to run the simulation a thousand times with random noise and see how much the results bounce around. That takes forever.

PRISM has a secret shortcut. It uses a mathematical tool called Fisher Information (think of it as a "Confidence Meter" built directly into the math).

Imagine you are trying to guess someone's age just by looking at their hand. If their hand changes shape wildly and predictably every year, you can guess their age very precisely (High Confidence).
If their hand looks the same at age 5 and age 6, you can't guess their age well (Low Confidence).
PRISM calculates this "guessability" instantly for every single point on the body. It tells you: "We are very sure about how the knee grows, but we are less sure about the soft tissue in the throat because that area varies a lot between people."

3. The "Time Traveler" (Intrinsic Time)

Sometimes, two kids are both 10 years old (chronological time), but one is a "late bloomer" and the other is an "early bloomer."

Chronological Time is what's on the calendar.
Intrinsic Time is their biological "maturity clock."

PRISM can look at a specific body part and say, "Even though this kid is 10, their airway looks like it belongs to a 9-year-old." It does this by reversing the growth map. It's like looking at a stretched rubber band and figuring out exactly how much it was stretched, then working backward to find the original state.

What Can PRISM Do? (Based strictly on the paper)

The authors tested PRISM on three things:

Predicting Growth: They showed it could accurately predict how a shape evolves over time, better than previous methods.
Spotting the "Odd Ones Out" (Anomaly Detection): This is a key application they tested. Imagine a child with a narrowed airway (a medical problem). PRISM can look at that specific narrowed spot and say, "This part of the airway looks like it belongs to a much younger child than the rest of this kid's body." It flags this as an anomaly without needing to be explicitly taught what a disease looks like.
Personalized Forecasts: If you know a child's current shape and their "biological age," PRISM can predict what their shape will look like in the future, keeping their unique growth pattern in mind.

The Bottom Line

PRISM is a new way for computers to understand 3D shapes (like organs or bones) as they change over time. It doesn't just give you a single "average" answer; it gives you a probabilistic map that shows the average and the uncertainty at every single point. It uses a clever math trick to calculate this uncertainty instantly, making it fast and useful for spotting developmental differences or medical issues in pediatric patients.

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

1. Problem Statement

Statistical shape modeling (SSM) aims to quantify biological variation by capturing the distribution of anatomical geometries. A critical challenge in healthcare research is modeling how anatomy evolves in response to continuous covariates (e.g., age $t$ ) while quantifying spatially varying uncertainties. Existing approaches face significant limitations:

Deterministic Models: Methods like NAISR achieve high-fidelity conditional generation but are fundamentally deterministic, providing point estimates without quantifying confidence or population variance.
Classical Statistical Atlases: Approaches based on diffeomorphic mapping (e.g., LDDMM) model variability in a parameter space (e.g., initial momenta). However, propagating this uncertainty to the image domain to obtain a pointwise map of aleatoric uncertainty on the anatomy is analytically intractable, typically requiring computationally intensive numerical resampling (e.g., Monte Carlo).
Gap: There is a lack of a framework that offers a theoretically grounded, closed-form formulation capable of directly mapping intrinsic biological ambiguity onto the anatomical space, enabling spatially resolved uncertainty quantification conditioned on covariates.

2. Methodology

PRISM (Probabilistic Interpretable Shape Modeling) bridges implicit neural representations with information geometry. It models shape evolution as a continuous, heteroscedastic Gaussian field conditioned on a covariate $t$ (e.g., chronological age).

2.1. Conditional Shape Distribution

The framework operates on shapes represented as displacement fields $\phi$ from a common template $T$ . For a template point $p$ and covariate $t$ , the displacement $d$ is modeled as:
$d | p, t \sim \mathcal{N}(\mu(p, t), \Sigma(p, t))$
where $\mu(p, t)$ is the mean developmental trajectory and $\Sigma(p, t)$ captures total population variability.

Architecture: A coordinate-based neural network $f(\cdot)$ parameterizes $\mu$ and the Cholesky factor $L$ of $\Sigma$ (where $\Sigma = LL^\top$ ).
Decomposition: The model separates covariate-driven terms from covariate-independent residuals to ensure identifiability.
Training: To avoid bias in mean estimation caused by variance-weighted gradients, PRISM employs a two-stage curriculum. It isolates gradient propagation by training the mean branch ( $\mu$ ) with an $\ell_1$ loss and the covariance branch ( $\Sigma$ ) with a negative log-likelihood (NLL) loss, using a warm-up phase for the mean before joint optimization.

2.2. Intrinsic Time Estimation

The framework introduces intrinsic time $\tau$ , a latent variable representing true developmental progression distinct from chronological time $t$ .

Inverse Encoder: An amortized inverse encoder $g(\cdot)$ is trained to map local observations (template point $p$ and displacement $d$ ) directly to an estimated intrinsic time $\hat{\tau}$ .
Training: The encoder is trained on synthetic triplets $(p, d, \tau)$ generated by the forward model, minimizing an $\ell_1$ loss. This allows for dense temporal inference across the anatomy without iterative test-time optimization.

2.3. Temporal Uncertainty Quantification

PRISM quantifies the uncertainty of intrinsic time estimation using the Fisher Information metric.

Theoretical Basis: Under the assumption that the population-average intrinsic time equals chronological time ( $E[\tau] = t$ ), the Cramér-Rao inequality states that the variance of the estimator is bounded by the inverse of the Fisher Information: $\text{Var}(\tau) \geq 1/I(p, t)$ .
Closed-Form Metric: For the heteroscedastic Gaussian model, the Fisher Information admits a closed-form expression decomposed into two terms:
1. $I_\mu$ (Mean Term): Information from the evolution of the mean trajectory.
2. $I_\Sigma$ (Covariance Term): Information from changes in population dispersion.
Selection: The authors retain only $I_\mu$ as the temporal uncertainty metric. They argue that while $I_\Sigma$ is statistically informative regarding population diversity, it is orthogonal to the goal of localizing an individual along the mean developmental trajectory.
Computation: The metric is computed analytically via automatic differentiation of the learned mean trajectory $\frac{\partial \mu}{\partial t}$ and the covariance $\Sigma^{-1}$ , avoiding sampling variance.

2.4. Applications

The unified framework supports:

Continuous Shape Trajectory: Generating normative population distributions at arbitrary spatiotemporal resolutions.
Intrinsic Time Estimation: Inferring global or local developmental stages.
Personalized Longitudinal Prediction: Forecasting future shapes by assuming a subject's temporal z-score remains constant.
Out-of-Distribution (OOD) Detection: Detecting anomalies (e.g., subglottic stenosis) by identifying regions where the local intrinsic time significantly lags behind the anatomy-wide median, normalized by local uncertainty.

3. Key Contributions

Conditional Probabilistic Implicit Field: A framework that jointly models the mean developmental trajectory and spatially varying population variability from cross-sectional data.
Closed-Form Fisher Information Metric: An analytically tractable method for efficient local uncertainty quantification via automatic differentiation, eliminating the need for Monte Carlo resampling.
Amortized Inverse Encoder: A mechanism to estimate intrinsic developmental time from local shapes without test-time optimization, enabling dense temporal inference.

4. Experimental Results

Experiments were conducted on three synthetic datasets (Starman G/L, ANNY) and one clinical dataset (Pediatric Airway).

Mean Trajectory Reconstruction: PRISM achieved competitive or superior performance in reconstruction error (Chamfer, Hausdorff, Earth Mover's Distance) compared to baselines A-SDF and NAISR. It effectively recovered global trends from sparse data.
Intrinsic Time Estimation:
- Global: PRISM's amortized inference was an order of magnitude faster than test-time optimization (TTO) baselines while maintaining comparable accuracy.
- Local: On the Starman (L) dataset, PRISM uniquely achieved near-perfect estimation of spatially varying intrinsic times (arms vs. legs), a capability baselines lacked.
Uncertainty Quantification: On synthetic data with ground truth, PRISM's predicted 95% confidence intervals aligned tightly with the true conditional distribution. On clinical airway data, the uncertainty bands varied spatially, correctly identifying soft tissue regions (e.g., tongue base) as having higher variability.
Personalized Prediction: PRISM outperformed baselines in longitudinal prediction tasks, demonstrating robust extrapolation beyond the training range.
OOD Detection: PRISM with local scoring significantly outperformed all global-time methods (AUC 0.875 vs. 0.605 for NAISR) in detecting pathological airway narrowing. This highlights the limitation of global approaches in capturing localized developmental deviations.

5. Significance and Claims

The paper claims PRISM is the first implicit neural representation framework for uncertainty-aware statistical shape analysis. Its significance lies in:

Interpretability: Providing spatially continuous, clinically meaningful uncertainty estimates directly on the anatomy.
Efficiency: Offering a closed-form solution for temporal uncertainty that is computationally efficient compared to numerical resampling.
Versatility: Demonstrating a unified framework capable of handling diverse tasks, from modeling shape evolution and personalized prediction to anomaly detection, within a single probabilistic model.
Clinical Utility: Successfully applying these concepts to pediatric airway analysis, where distinguishing physiological variation from pathology is critical.

The authors note limitations, including the current reliance on a single scalar covariate and the need for further validation in long-term forecasting for degenerative conditions. They emphasize that clinical translation requires external validation on target populations and caution that predictions should be interpreted carefully when individual variation approaches covariate-driven variation.