Equivariant bootstrapping for uncertainty quantification in imaging inverse problems

Marcelo Pereyra, Julián Tachella

Abstract

Scientific imaging problems are often severely ill-posed and hence have significant intrinsic uncertainty. Accurately quantifying the uncertainty in the solutions to such problems is therefore critical for the rigorous interpretation of experimental results as well as for reliably using the reconstructed images as scientific evidence. Unfortunately, existing imaging methods are unable to quantify the uncertainty in the reconstructed images in a way that is robust to experiment replications. This paper presents a new uncertainty quantification methodology based on an equivariant formulation of the parametric bootstrap algorithm that leverages symmetries and invariance properties commonly encountered in imaging problems. Additionally, the proposed methodology is general and can be easily applied with any image reconstruction technique, including unsupervised training strategies that can be trained from observed data alone, thus enabling uncertainty quantification in situations where there is no ground truth data available. We demonstrate the proposed approach with a series of experiments and comparisons with alternative state-of-the-art uncertainty quantification strategies. In all our experiments, the proposed equivariant bootstrap delivers remarkably accurate high-dimensional confidence regions and outperforms the competing approaches in terms of estimation accuracy, uncertainty quantification accuracy, and computing time. These empirical findings are supported by a detailed theoretical analysis of equivariant bootstrap for linear estimators.

Introduction

In scientific fields like medicine or astronomy, researchers often face inverse problems: reconstructing a clear image from messy, incomplete, or noisy data, such as CT scans or blurry telescope captures. While modern AI models excel at generating high-quality reconstructions, they rarely provide a measure of confidence in their output.

Current methods for Uncertainty Quantification (UQ) face significant hurdles. Bayesian methods, such as those utilizing Diffusion models, are computationally expensive, often requiring hundreds of neural network passes. Conversely, standard statistical tools like the "Naive" Bootstrap fail in imaging because they are "blind" to the nullspace—the parts of the image the sensor did not capture. This leads to overconfidence and underestimated error margins. This work introduces the Equivariant Bootstrap to provide a mathematically rigorous, fast, and reliable UQ framework.

Methods: The Equivariant Bootstrap

The core innovation is the Equivariant Bootstrap, which leverages natural symmetries (like rotations and shifts) to probe the estimator's bias. In a standard bootstrap, one takes a reconstructed image, simulates sensor data, and reconstructs it again. However, the AI often repeats its previous errors on the same image, creating a false sense of certainty.

The Equivariant version introduces a transformation step:

  1. Transform: Apply a random transformation $g$ (e.g., rotation) to the initial reconstruction $\hat{x}$.
  2. Measure & Reconstruct: Simulate the measurement process on this transformed image and use the AI to reconstruct it.
  3. Invert: Undo the transformation to obtain a "bootstrap sample" $\tilde{x}^{(i)}$.
The Equivariant Bootstrap pipeline: applying transformations to probe uncertainty.

By rotating the image before measurement, the AI is forced to reconstruct features from a different "angle" relative to the sensor's fixed geometry. This process reveals the true underlying uncertainty by averaging out the estimator's bias.

Experimental Results

The method was evaluated across three challenges: Compressed Sensing (MNIST), Inpainting (DIV2K), and Sparse-Angle Tomography (LIDC-IDRI).

Accuracy and Coverage

A key metric for UQ is empirical coverage: how often the true image falls within the predicted confidence interval. In high-dimensional tasks like DIV2K inpainting, the proposed method achieved nearly ideal coverage, closely following the diagonal in coverage plots. In contrast, competing methods like Unadjusted Langevin Algorithm (ULA) or Diffusion models (DDRM, diffPIR) either significantly underestimated or overestimated uncertainty.

Empirical coverage plots across different imaging tasks, comparing the proposed method to state-of-the-art baselines.

Performance and Speed

In the Tomography task, the equivariant bootstrap achieved a PSNR of 41.03 dB, outperforming ULA (35.85 dB) and DiffPIR (37.02 dB). Crucially, the method is orders of magnitude faster than Bayesian MCMC methods. While Diffusion models might require 100 Neural Function Evaluations (NFEs) per sample, the bootstrap method requires only 1 NFE per sample, delivering results in seconds rather than minutes or hours.

Conclusion

The Equivariant Bootstrap represents a significant advancement in UQ for imaging. By using symmetries as a probe for the nullspace, it provides reliable "error bars" even when trained without ground truth data. Future work includes automatic calibration for selecting optimal transformations and extending the framework to other data types with inherent symmetries, such as time series or graphs.

Summary updated

Abstract

Scientific imaging problems are often severely ill-posed and hence have significant intrinsic uncertainty. Accurately quantifying the uncertainty in the solutions to such problems is therefore critical for the rigorous interpretation of experimental results as well as for reliably using the reconstructed images as scientific evidence. Unfortunately, existing imaging methods are unable to quantify the uncertainty in the reconstructed images in a way that is robust to experiment replications. This paper presents a new uncertainty quantification methodology based on an equivariant formulation of the parametric bootstrap algorithm that leverages symmetries and invariance properties commonly encountered in imaging problems. Additionally, the proposed methodology is general and can be easily applied with any image reconstruction technique, including unsupervised training strategies that can be trained from observed data alone, thus enabling uncertainty quantification in situations where there is no ground truth data available. We demonstrate the proposed approach with a series of experiments and comparisons with alternative state-of-the-art uncertainty quantification strategies. In all our experiments, the proposed equivariant bootstrap delivers remarkably accurate high-dimensional confidence regions and outperforms the competing approaches in terms of estimation accuracy, uncertainty quantification accuracy, and computing time. These empirical findings are supported by a detailed theoretical analysis of equivariant bootstrap for linear estimators.
BibTeX
@inproceedings{pereyra2024equivariant,
  author = {Marcelo Pereyra and Julián Tachella},
  title = {{Equivariant bootstrapping for uncertainty quantification in imaging inverse problems}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2024},
  url = {https://openreview.net/forum?id=86x3OWfKJk},
}
Figure