Source-linked AI summary
MUMINS: Metadata-conditioned Uncertainty-aware Medical Image Next-state Synthesis
Anna Oliveras, Roger Marí, Rafael Redondo, Oriol Guardià, Cynthia Ifeyinwa Ugwu, Ana Tost, Bhalaji Nagarajan, Carolina Migliorelli, Vicent Ribas, Petia Radeva
TL;DR
Forecasting patient-specific anatomical change is difficult because evolution is subtle, stochastic, and difficult to represent with deterministic models or costly multi-pass uncertainty estimation. MUMINS jointly diffuses baseline and residual information while predicting a spatial uncertainty map in one reverse process, using metadata conditioning and baseline reinjection. After dataset-specific retraining, it matches or outperforms domain-specific methods on lung CT and brain MRI, with limitations in variance-propagation assumptions and uncertainty localization.
Problem
Anatomical progression forecasting must represent subtle, patient-specific, stochastic changes while preserving static anatomy and estimating uncertainty efficiently.
Source: §1 · p. 2Method
MUMINS jointly denoises a baseline scan and follow-up residual, reinjects the baseline during sampling, conditions on metadata, and predicts voxel-wise uncertainty in one reverse process.
Source: §1 · p. 2, §3.3 · p. 4Results
Takeaways & Limitations
A universal architecture can support follow-up synthesis and inexpensive single-pass uncertainty estimation across lung CT and brain MRI datasets.
Source: §1 · p. 2, §5 · p. 9Inference-time variance propagation assumes mutually independent diagonal components and omits the denoiser Jacobian because 3D backward passes are infeasible at scale.
Source: §3.6 · p. 5
Abstract
from arXiv · showhide
Forecasting anatomical changes such as tumor growth and neurodegeneration is a challenging generative vision task. Morphological evolution is subtle relative to static anatomy, highly patient-specific, and inherently stochastic. Existing methods struggle with several issues: deterministic networks ignore biological stochasticity, while standard diffusion models require computationally prohibitive multi-pass sampling to quantify uncertainty. We propose MUMINS (Metadata-conditioned Uncertainty-aware Medical Image Next-state Synthesis), an efficient diffusion framework that jointly diffuses a baseline scan and its follow-up residual, summed to synthesize the follow-up scan, while concurrently predicting a spatial uncertainty map, in a single reverse diffusion process. Conditioned on the time interval and relevant metadata, it preserves fine-grained anatomy by dynamically re-injecting the baseline as a soft anchor at every denoising step, and a negative-log-likelihood head learns the uncertainty map to explicitly flag error-prone regions. Designed without organ-specific heuristics, the same architecture is reused across anatomies via separate, dataset-specific retraining. Extensive evaluations demonstrate that dataset-specific retraining of MUMINS matches or outperforms dedicated, domain-specific state-of-the-art methods on lung CT (PNG) and brain MRI (OASIS-3). Project page: https://github.com/aolivtous/MUMINS.
1. Introduction
MUMINS addresses the stochastic, patient-specific difficulty of forecasting anatomical change while preserving baseline anatomy and estimating spatial uncertainty. Its universal diffusion design matches or outperforms domain-specific methods after dataset-specific retraining.
- The clinical motivation spans lung-nodule assessment and longitudinal neurodegeneration tracking, where future scans at arbitrary intervals support prognostic decisions.Source: §1 · p. 2
- Forecasting is difficult because morphological change is small relative to static anatomy, patient-specific, and inherently stochastic, requiring plausible changes alongside preserved baseline details.Source: §1 · p. 2
- Existing methods remain limited by deterministic single-future predictions, organ-specific components, and costly multi-pass uncertainty estimation.Source: §1 · p. 2
- MUMINS jointly denoises the baseline and baseline-to-follow-up residual, reinjecting the baseline during inference to preserve anatomy while synthesizing morphological change.Source: §1 · p. 2
- A single reverse-diffusion process predicts voxel-wise uncertainty through likelihood-based learning, avoiding Monte Carlo’s repeated inferences per sample.Source: §1 · p. 2
- MUMINS matches or outperforms domain-specific state-of-the-art methods on PNG lung CT and OASIS-3 brain MRI after dataset-specific retraining.Source: §1 · p. 2
2. Related Work
Longitudinal medical image progression methods seek to forecast patient-specific anatomical change and quantify uncertainty. Existing deterministic and diffusion-based approaches leave gaps in modeling plausible follow-ups and efficiently estimating uncertainty.
- Longitudinal progression forecasting targets patient-specific anatomical change from baseline scans, including pulmonary nodule evolution and neurodegenerative brain progression.Source: §2 · p. 2
Deterministic and registration-based methods
- Deterministic and registration-based models learn fixed baseline-to-follow-up transitions, including voxel warping and intensity refinement for lung nodules.Source: §2.1 · p. 2
Diffusion-based progression
- Diffusion models can generate high-fidelity follow-up images, but recent progression methods remain strongly anatomy-, disease-, or modality-specific.Source: §2.1 · p. 3
- Monte Carlo uncertainty estimation requires N × T network evaluations per volume, while BayesDiff still resamples weights at every step and retains the multi-pass bottleneck.Source: §2.2 · p. 3
3. Methodology
MUMINS jointly models baseline anatomy and longitudinal residuals in a metadata-conditioned 3D diffusion process, producing follow-up scans and voxel-wise uncertainty. Baseline re-injection anchors synthesis to patient anatomy, while variance propagation and NLL training make single-pass uncertainty estimation tractable.
Longitudinal formulation
- MUMINS jointly denoises the baseline and follow-up residual, then adds the denoised residual to the baseline to synthesize the future scan.The residual formulation separates static anatomy from longitudinal change.Source: §3.1 · p. 3, §3 · p. 3, §3.5 · p. 4
Diffusion model
- A 3D U-Net predicts diffusion noise and per-voxel standard deviation for the joint baseline-residual state.The variance head supports uncertainty learning alongside noise prediction.Source: §3.2 · p. 3, §3.4 · p. 4
Temporal and metadata conditioning
- Conditioning combines the inter-scan interval, baseline age, and per-axis blur changes, with Δb = 0 as the information-respecting forecasting setting.The oracle Δb = GT setting is reported only as a diagnostic ceiling.Source: §3.3 · p. 4
Hybrid training objective
- The hybrid objective combines joint noise reconstruction with an NLL uncertainty term, using stop-gradient decoupling and λ_NLL = 0.01.This prevents uncertainty calibration from reducing NLL by inflating predicted standard deviations at the expense of denoising.Source: §3.4 · p. 4
Baseline injection
- At every reverse step, MUMINS overwrites the baseline channel with a forward-marginal sample, using one fixed noise realization to maintain a coherent anatomical trajectory.The residual is thereby guided by temporally consistent baseline context; single-trajectory injection is preferred to per-step resampling in the cited ablation.Source: §3.5 · p. 4
Uncertainty propagation
- Variance propagation uses diagonal covariance and J_t = 0 approximations, with delayed activation selected by NLL to avoid monotonically accumulating over-dispersed uncertainty.Stochastic sampler noise is required for both calibrated uncertainty and reconstruction quality in the reported ablation.Source: §3.6 · pp. 4–5
4. Experiments and Results
MUMINS is evaluated on lung CT and brain MRI for synthesis quality, uncertainty estimation, qualitative fidelity, and component contributions. It performs strongly against organ-specific baselines while its uncertainty estimates provide useful error localization and calibration at substantially lower sampling cost.
Experimental setup
- The experiments use PNG lung CT and OASIS-3 brain MRI, with metrics averaged over longitudinal pairs and synthesis assessed by MAE, PSNR, and SSIM.Uncertainty is evaluated against voxel-wise absolute error for both whole-image and relevant ROI settings.Source: §4.1 · p. 5, §4.3 · pp. 5–6, §4.4 · p. 6
Synthesis results
- MUMINS achieves the best aggregate whole-image lung metrics and leads aggregate nodule-ROI PSNR, while brain performance is ahead of TADM-3D and BrLP on whole-image metrics.On PNG, whole-image PSNR is 23.53 versus 22.85 dB for NGP-Net, and nodule-ROI PSNR is 24.89 dB; on OASIS-3, MUMINS is ahead of TADM-3D and BrLP on all three whole-image metrics.Source: §4.4 · pp. 6–7, Table 1 · p. 7
Qualitative results
- Qualitatively, MUMINS produces sparse, low-amplitude lung errors and clean brain reconstructions, whereas competing methods show broader errors or blurrier follow-ups.On lung CT, little error appears over the nodule; on brain MRI, MUMINS is described as cleanest overall while BrLP is visibly blurrier.Source: Figure 2 · p. 6, §4.4 · p. 7
Uncertainty evaluation
- MUMINS uncertainty localizes lung error with ρ = 0.61 and PICP95 = 0.91, while brain ROI uncertainty remains better calibrated than the 20-seed MC reference.For brain ROI, single-pass uncertainty has ρ = 0.26 versus 0.32 for MC, but PICP95 = 0.97 versus 0.59; cross-estimator agreement is 0.79 for lung and 0.65 for brain ROI.Source: §4.4 · p. 8, Table 2 · p. 8
- The single-pass sparsification curve approaches oracle behavior as high-uncertainty voxels are removed, though it remains above the 20-seed MC curve.The higher single-pass AUSE is consistent with weaker sparsification than MC, while both curves show monotonically decreasing MAE.Source: §4.4 · p. 8, Figure 4 · p. 8
Ablation studies
- Baseline reinjection produces the largest ablation gains, increasing PSNR from 21.00 to 23.73 dB and SSIM from 56.27% to 72.38%.Residual prediction reduces PSNR standard deviation from 4.09 to 1.65 dB, blur conditioning raises PSNR from 17.40 to 20.69 dB, and uncertainty prediction adds a further gain to 21.00 dB.Source: §4.5 · p. 8, Table 3 · p. 8
5. Conclusions
MUMINS synthesizes follow-up scans and voxel-wise uncertainty in one reverse diffusion process, using joint baseline-residual modeling and baseline re-injection to preserve anatomy. It matches or outperforms domain-specific baselines on lung CT and brain MRI, while retaining limitations in brain localization and cross-horizon consistency.
- MUMINS matches or outperforms domain-specific baselines on lung CT and brain MRI, with single-pass uncertainty correlating cheaply with the Monte Carlo reference.The framework uses no organ-specific components, but brain uncertainty localization trails Monte Carlo on near-saturated targets.Source: §5 · p. 9
- MUMINS synthesizes a follow-up scan and voxel-wise uncertainty map in a single reverse diffusion process.Its joint baseline-residual state and per-step baseline re-injection preserve fine anatomy while generating plausible change.Source: §5 · p. 9
- The likelihood-based variance head provides uncertainty without Monte Carlo’s costlier repeated sampling.This enables single-pass uncertainty estimation while retaining correlation with the Monte Carlo reference.Source: §5 · p. 9
- Predictions queried at different future intervals from the same baseline are generated independently, without explicit cross-horizon consistency.The authors identify this as a remaining limitation and propose extending the framework to further organs and modalities.Source: §5 · p. 9