Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Uncertainty Estimation for Pretrained Medical Image Registration Models via Transformation Equivariance

NeurIPS 2026 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Priors and functional variational uncertainty · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes an inference-time, model-agnostic uncertainty estimator for pretrained deformable image registration networks by perturbing the source image with random spatial transforms, undoing the transform on the predicted deformation, and taking the per-voxel variance, with a theoretical decomposition linking it to registration error.

Problem

Pretrained deep registration networks give deformation fields without indicating where or when they are unreliable, and existing uncertainty methods (Bayesian, ensembles, MC-dropout) require retraining or architecture changes.

Approach

Uses transformation equivariance: perturb the source image by random transforms, compose each predicted field with the inverse perturbation, and treat the sample variance of the composed fields as uncertainty; a first/second-order analysis decomposes this variance into intrinsic spread and bias jitter linked to registration error.

Key results

• Proposed uncertainty maps show strong spatial correlation with MC-dropout uncertainty on TransMorph-Bayes IXI test cases (single cases and dataset mean reported).
• For uniGradICON, uncertainty correlates positively with voxel registration error across translation, affine, deformation and real settings, strong for linear ground truth and moderate for nonlinear ground truth.
• Inference overhead for uncertainty with N=50 perturbations is about 12 s for transformation perturbations and 27.7 s for deformation perturbations on an RTX 4500 Ada GPU, versus 0.20 s for the uniGradICON backbone.

Surprising findings

• For SynthMorph on translation and affine ground truth, uncertainty correlation with error drops sharply, attributed to systematic bias that variance does not capture.

Limitations

• Theory relies only on first- and second-order moments of the error model; higher-order statistics unexplored.
• Uncertainty does not disentangle sources of uncertainty as Bayesian methods do.
• Only pretrained deep registration networks evaluated; not applied to optimization-based registration.

Experimental setup

Datasets: IXI brain MRI ACDC cardiac MRI Learn2Reg abdomen CT curated brain MRI (11 public sources, 220 volumes) BraTS-Reg Learn2Reg ThoraxCBCT

Benchmarks: IXI brain MRI ACDC cardiac MRI Learn2Reg abdomen CT curated brain MRI BraTS-Reg Learn2Reg ThoraxCBCT

Metrics: Pearson correlation Spearman correlation AURC mean Euclidean registration error (mError) runtime GPU memory

Baselines: MC-dropout TransMorph-Bayes

Builds on: MC-dropout test-time augmentation TransMorph

Closest papers by meaning

• Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging (ICLR 2025 · similarity 0.91)
• CARL: A Framework for Equivariant Image Registration (CVPR 2025 · similarity 0.90)
• A Variational Bayesian Method for Similarity Learning in Non-Rigid Image Registration (CVPR 2022 · similarity 0.87)
• Deep Learning in Medical Image Registration: Magic or Mirage? (NeurIPS 2024 · similarity 0.84)
• Rate-In: Information-Driven Adaptive Dropout Rates for Improved Inference-Time Uncertainty Estimation (CVPR 2025 · similarity 0.84)
• Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration (CVPR 2024 · similarity 0.84)