Denoising diffusion networks for normative modeling in neuroimaging

Normative modeling estimates reference distributions of biological measures conditional on covariates, enabling centiles and clinically interpretable deviation scores to be derived. Most neuroimaging pipelines fit one model per imaging-derived phenotype (IDP), which scales well but discards multivariate dependence that may encode coordinated patterns. We propose denoising diffusion probabilistic models (DDPMs) as a unified conditional density estimator for tabular IDPs, from which univariate centiles and deviation scores are derived by sampling. We utilise two denoiser backbones: (i) a feature-wise linear modulation (FiLM) conditioned multilayer perceptron (MLP) and (ii) a tabular transformer with feature self-attention and intersample attention (SAINT), conditioning covariates through learned embeddings. We evaluate on a synthetic benchmark with heteroscedastic and multimodal age effects and on UK Biobank FreeSurfer phenotypes, scaling from dimension of 2 to 200. Our evaluation suite includes centile calibration (absolute centile error, empirical coverage, and the probability integral transform), distributional fidelity (Kolmogorov-Smirnov tests), multivariate dependence diagnostics, and nearest-neighbour memorisation analysis. For low dimensions, diffusion models deliver well-calibrated per-IDP outputs comparable to traditional baselines while jointly modeling realistic dependence structure. At higher dimensions, the transformer backbone remains substantially better calibrated than the MLP and better preserves higher-order dependence, enabling scalable joint normative models that remain compatible with standard per-IDP pipelines. These results support diffusion-based normative modeling as a practical route to calibrated multivariate deviation profiles in neuroimaging.

💡 Research Summary

This paper introduces a novel approach to normative modeling of neuroimaging-derived phenotypes (IDPs) by leveraging denoising diffusion probabilistic models (DDPMs) as conditional density estimators. Traditional pipelines fit a separate model for each IDP, which scales well but discards multivariate dependencies that may encode coordinated disease patterns. The authors propose to model the joint conditional distribution p(y | c) of all IDPs y given covariates c (age, sex) using a diffusion process, from which univariate centiles and multivariate deviation scores can be derived by sampling.

Two denoiser backbones are evaluated: (i) a FiLM‑conditioned multilayer perceptron (MLP) that modulates each layer with learned linear transformations of the covariates, and (ii) a SAINT‑style tabular transformer that applies self‑attention across features and inter‑sample attention across mini‑batches, conditioning covariates via learned embeddings. The diffusion model follows the standard forward noising schedule q(y_t | y_{t‑1}) = N(√{1‑β_t} y_{t‑1}, β_t I) and learns to predict the added noise ε_θ(y_t, t, c). Training minimizes the usual L2 noise‑prediction loss with timesteps sampled uniformly.

Evaluation is performed on two datasets. A synthetic benchmark (N = 47 000) is constructed with heteroscedastic, nonlinear, and bimodal age effects to test calibration under challenging distributions. The real‑world dataset consists of UK Biobank FreeSurfer IDPs. For low‑dimensional experiments (D ≤ 20) a clinically relevant set of 20 neurodegeneration‑related phenotypes is used; for higher dimensions (up to D = 200) random additional IDPs are added to stress scalability while preserving the core set.

A comprehensive evaluation suite is introduced:

1. Centile calibration – absolute centile error, empirical coverage, and probability integral transform (PIT) histograms.
2. Distributional fidelity – Kolmogorov–Smirnov (KS) tests comparing sampled and empirical marginal distributions.
3. Multivariate dependence diagnostics – correlation matrix recovery, higher‑order dependence measures, and visualisation after dimensionality reduction.
4. Nearest‑neighbour memorisation analysis – assessing whether the model merely reproduces training points.

Results show that in low dimensions diffusion models achieve calibration comparable to traditional baselines (GAMLSS, hierarchical Bayesian regression, Gaussian processes) while simultaneously modeling realistic dependence across IDPs. In higher dimensions the FiLM‑MLP degrades sharply: centile errors increase, KS p‑values drop, and the recovered correlation structure weakens. Conversely, the transformer‑backed diffusion model remains well‑calibrated (absolute centile error ≈ 0.03–0.05, coverage ≈ 95 %), preserves higher‑order dependencies (correlation recovery R ≈ 0.85), and exhibits no harmful memorisation.

The study demonstrates that (a) DDPMs can naturally capture heteroscedasticity, multimodality, and non‑linear covariate effects in tabular neuroimaging data; (b) attention‑based transformers provide the expressive power needed for high‑dimensional joint modeling; and (c) a diffusion‑based normative framework yields multivariate deviation profiles that are compatible with existing per‑IDP pipelines yet richer in information.

The authors discuss future directions, including faster sampling schemes (e.g., DDIM, accelerated reverse processes), richer conditioning mechanisms (meta‑learning, hierarchical embeddings), and methods for improving clinical interpretability (feature‑wise contribution scores, conditional sharpening). Overall, the work positions diffusion models, particularly when paired with transformer denoisers, as a practical and scalable solution for calibrated multivariate normative modeling in large neuroimaging cohorts.