Introduction
Scientific imaging often involves solving inverse problems—reconstructing high-fidelity images from noisy or incomplete measurements. These problems are typically ill-posed, where multiple reconstructions can satisfy the observed data. While deep learning has revolutionized reconstruction, quantifying the uncertainty of these "guesses" remains a challenge. Existing methods like Bayesian MCMC or diffusion models are computationally prohibitive, while standard bootstrapping often underestimates error due to inherent model bias.
The Equivariant Bootstrap
To address these limitations, we introduce Equivariant Bootstrapping. This method leverages the natural symmetries (e.g., rotations, translations) inherent in many image datasets. Unlike the standard parametric bootstrap, which simulates new data directly from a reconstructed image $\hat{x}$, the equivariant approach introduces a random transformation $g$ from a symmetry group $G$ before simulation.
The algorithm follows three main steps:
- Apply a random transformation $g$ (e.g., a 90-degree rotation) to the initial reconstruction $\hat{x}$.
- Simulate a new measurement from this transformed image and reconstruct a new sample using the original model.
- Apply the inverse transformation $g^{-1}$ to the result to align it with the original coordinate frame.
By "averaging out" reconstruction biases through these transformations—a process related to Reynolds averaging in group theory—the method effectively explores the nullspace of the imaging operator. This allows for accurate uncertainty quantification without requiring ground truth labels. Crucially, this method is model-agnostic and can be wrapped around any existing reconstruction network.
Experimental Results
The method was validated across three distinct imaging challenges: MNIST compressed sensing, DIV2K inpainting, and sparse-angle tomography (LIDC-IDRI).
- Accuracy: Coverage plots demonstrated that Equivariant Bootstrapping provides highly reliable confidence intervals, significantly outperforming over-confident diffusion models. It ensures that a 90% confidence interval actually contains the truth approximately 90% of the time.
- Efficiency: The method requires only one Neural Function Evaluation (NFE) per sample, making it up to 100x faster than iterative Bayesian samplers or diffusion models which often require 100+ NFEs.
- Unsupervised Applicability: It successfully quantified uncertainty even for models trained without ground truth, a critical requirement for many real-world scientific sensors where "truth" is never known.
Conclusion
Equivariant Bootstrapping offers a scalable, model-agnostic framework for uncertainty quantification in imaging. By utilizing known symmetries, it breaks the bias of standard bootstrapping and provides a computationally efficient alternative to state-of-the-art generative models. Future work will focus on automating the selection and calibration of symmetry groups for diverse imaging modalities.
Summary updated