Self-supervised conformal prediction for uncertainty quantification in Poisson imaging problems

Bernardin Tamo Amougou, Marcelo Pereyra, Barbara Pascal · 2025

Read Self-supervised conformal prediction for uncertainty quantification in Poisson imaging problems by Bernardin Tamo Amougou, Marcelo Pereyra, Barbara Pascal (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-19194.txt.

Introduction

Bernardin Tamo Amougou and co-authors put honest error bars on restored images without ever seeing the truth. Restoration is ill-posed, so a cleaned image hides real uncertainty, yet conformal prediction, the rigorous way to size it, normally needs ground truth for calibration. Poisson noise, the grain of photon-counting cameras, is their setting.

Their trick swaps ground truth for the Poisson unbiased risk estimator, letting the method calibrate itself from the measurements alone. The result applies to any ill-conditioned Poisson imaging problem and pairs naturally with self-supervised restorers trained on raw data. On denoising and deblurring tests it matched supervised conformal methods that cheat with true images. The authors offer it as uncertainty quantification for fields where truth is exactly what is missing.

Catalog record

CREATORBernardin Tamo Amougou, Marcelo Pereyra, Barbara Pascal
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-19194

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Self-supervised conformal prediction for uncertainty quantification in Poisson imaging problems Thanks: This work was supported by French National Research Agency (ANR) (ANR-23-CE48-0009, “ OptiMoCSI "); and by UKRI Engineering and Physical Sciences Research Council (EPSRC) (EP/V006134/1, EP/Z534481/1). Bernardin Tamo Amougou, Marcelo Pereyra, Barbara Pascal Abstract Image restoration problems are often ill-posed, leading to significant uncertainty in reconstructed images. Accurately quantifying this uncertainty is essential for the reliable interpretation of reconstructed images. However, image restoration methods often lack uncertainty quantification capabilities. Conformal prediction offers a rigorous framework to augment image restoration methods with accurate uncertainty quantification estimates, but it typically requires abundant ground truth data for calibration. This paper presents a self-supervised conformal prediction method for Poisson imaging problems which leverages Poisson Unbiased Risk Estimator to eliminate the need for ground truth data. The resulting self-calibrating conformal prediction approach is applicable to any Poisson linear imaging problem that is ill-conditioned, and is particularly effective when combined with modern self-supervised image restoration techniques trained directly on measurement data. The proposed method is demonstrated through numerical experiments on image denoising and deblurring; its performance are comparable to supervised conformal prediction methods relying on ground truth data. I Introduction Poisson imaging problems are widely encountered across various scientific and engineering disciplines, with notable prevalence in fields such as astronomy and microscopy [ 1 , 2 ] ; common examples include Poisson image denoising, deblurring and computed tomography [ 3 , 4 , 5 ] . Such problems are often not well posed, meaning that there exists a wide range of possible solutions that are in agreement with the observed data. Accounting for this inherent uncertainty is essential when using the reconstructed images as evidence for science or critical decision-making [ 6 ] . Unfortunately, most methods for solving Poisson imaging problems currently available are unable to accurately quantify the uncertainty in the delivered solutions. Bayesian statistical inference strategies, which rely on prior knowledge about the likely solutions to the problem as encoded by a probability distribution, are the predominant approach to uncertainty quantification in imaging problems (see, e.g., [ 3 , 7 , 8 , 9 ] ). Bayesian imaging techniques can deliver a wide range of inferences, and the most accurate solutions to Poisson imaging problems are currently obtained by using Bayesian imaging methods combining ideas from optimisation, deep learning and stochastic sampling [ 9 ] . However, even the most advanced Bayesian imaging strategies currently available struggle to provide accurate uncertainty quantification on structures larger than a few pixels in size [ 10 ] . This has stimulated research on other frameworks for performing uncertainty quantification in imaging inverse problems. Namely, bootstrapping and conformal prediction have recently emerged as promising frameworks for delivering accurate uncertainty quantification for large image structures. With regards to bootstrapping, we note the recent equivariant bootstrapping method [ 11 ] which yields excellent confidence regions, even for large image structures. Equivariant bootstrapping is a statistical resampling strategy that takes advantage of known symmetries of the problem to significantly reduce the bias inherent to synthetic resampling. It does not require ground truth data for calibration, and it can be therefore applied to many quantitative and scientific imaging problems for which obtaining reliable ground truth data is difficult or impossible. Equivariant bootstrapping is effective in problems that are severely ill-posed, for example in compressive sensing, inpainting, limited angle tomography or radio-interferometry [ 11 , 12 ] . Conversely, it often performs poorly for problems which are potentially severely ill-conditioned but not ill-posed, as this prevents leveraging symmetries to remove the bootstrapping bias [ 11 ] . In contrast, conformal prediction is a fully data-driven strategy for constructing regions with almost exact marginal coverage [ 13 ] . Unlike equivariant bootstrapping, conformal prediction is agnostic to both the image observation model and our prior knowledge about likely solutions. However, in its original form, conformal prediction requires abundant ground truth data for calibration, which is a main bottleneck to implementation for quantitative and scientific imaging problems where ground truth data is often scarce or unavailable. The need for ground truth can be partially mitigated by combining conformal prediction with Bayesian inference and by focusing on pixels of a single image rather than regions, as proposed in [ 14 ] . Alternatively, the recent method [ 15 ] fully bypasses the need for ground truth data by focusing on Gaussian imaging problems and leveraging Stein’s Unbiased Risk Estimator [ 16 ] to perform conformal prediction directly from the observed measurement data, in a self-supervised manner. The experiments reported in [ 15 ] show that, for non-blind linear Gaussian imaging problems, self-supervised conformal prediction is as accurate as supervised conformal prediction. There are many important imaging problems for which the Gaussian noise assumption underpinning [ 15 ] is not sufficiently accurate to deliver meaningful inferences. In particular, imaging sensors often exhibit so-called shot noise, a form of signal-dependent measurement error that exhibits Poisson statistics. With quantitive and scientific Poisson imaging problems in mind, this paper proposes a generalisation of the self-supervised conformal prediction framework [ 15 ] for high-dimensional linear inverse problems involving Poisson noise. The remainder of the paper is organised as follows. The principles of conformal prediction for inverse problems involving Poisson noise are introduced in Section II . Then, the proposed self-supervised conformal prediction framework is detailed in Section III . This framework is demonstrated on Poisson image denoising and deblurring experiments in Section IV . Concluding remarks and future research directions are provided in Section V . II Conformal Prediction under Poisson models In the context of image estimation, conformal prediction aims to identify a set of statistically likely values for an unknown image of interest x^{\star} from some noisy measurements y\in\mathbb{R}^{m} of x^{\star} . Assuming x^{\star} and y are realizations of random variables X and Y , for a confidence level \alpha\in(0,1) , we seek a region C(Y)\subset\mathbb{R}^{n} verifying \mathbb{P}_{(X,Y)}\left(X\in C(Y)\right)\geq 1-\alpha\, (1) where the probability is taken w.r.t. the joint distribution of (X,Y) . Conformal prediction is a general framework leveraging a non-conformity measure s:\mathbb{R}^{n}\times\mathbb{R}^{m}\to\mathbb{R} , which reflects how plausible is an image x given measurements y , to provide such confidence sets [ 13 ] . Once chosen s , the set C(y) is constructed by computing the top (1-\alpha) -quantile q_{\alpha} of the statistic s(X,Y) , i.e., \displaystyle C(y):=\{x\in\mathbb{R}^{n}\,|\,s(x,y)\leq q_{\alpha}\}\quad\text{for all }y\in\mathbb{R}^{m}. (2) which by construction set satisfies: \displaystyle\mathbb{P}_{(X,Y)}\big(X\in C(Y)\big)=\mathbb{P}_{(X,Y)}\big(s(X,Y)\leq q_{\alpha}\big)\geq 1-\alpha, (3) for any confidence level \alpha\in(0,1) . Provided that s has been suitably designed and that enough samples \{x_{i},y_{i}\}_{i=1}^{M} under the joint distribution of (X,Y) are available to calibrate q_{\alpha} , then ( 2 ) yields regions C(y) that contain x^{\star} with the prescribed probability. In this paper, we consider a linear Poisson, i.e., shot noise, observation model, in which X and Y are related by (Y\mid X=x^{\star})\sim\mathcal{P}\!\Bigl(\gamma Ax^{\star}\Bigr)\,, where \mathcal{P} denotes the Poisson distribution, A is a linear measurement operator, and \gamma>0 is related to the strength of the shot noise (the larger \gamma , the easier the image restoration problem). We henceforth assume A and \gamma are known, and that A is full rank but possibly severely ill-conditioned. To design C(y) , we assume the availability of some estimator of x^{\star} ,

Read the complete text