Biologists increasingly run image-restoration software they cannot audit. Tools such as deep-learning denoisers and deconvolution programs (which undo optical blur) sharpen fluorescence images. When the microscope, sample, or laser power differs from what the tool assumed, these tools can erase real structure or invent structure that is not there. Nothing on screen warns the user. This proposal gives every reconstruction a check a non-expert can read, called a residual certificate. The residual is what is left after subtracting the predicted clean image, pushed back through a model of the microscope, from the raw image. If the physics model and the noise model are right, that leftover should look like featureless, unit-variance noise everywhere. Where it does not, the reconstruction cannot be trusted. The same test can also choose the one parameter users struggle with most: how strongly to smooth.
The core claim has two parts. First, a local whiteness test on noise-normalized residuals can select the smoothing strength of a variational deconvolution solver close to the best possible choice, under real mixed shot-plus-read noise. Second, the same statistic, computed patch by patch, gives a spatial trust map that flags a wrong blur model or a noise-level shift. The test is model-agnostic: it only needs the output image, the raw image, and the forward model, so it applies equally to classical solvers and to learned networks already distributed through the BioImage Model Zoo. The accessibility gain is not a new interface. It is a replacement for a mathematical knob with one automatic choice and one picture a biologist can interpret.
Why this, and not an agent
The obvious idea at this intersection is an LLM agent that tunes deconvolution through conversation. It would transplant CellAgent, a planner–executor–evaluator system that judges single-cell plots with GPT-4o, and AutoMat, an agent that chains physics tools to rebuild crystal structures from electron micrographs. Both work because a closed-loop evaluator catches bad intermediate results.
In optical deconvolution, however, the right evaluator already exists in mathematical form. The forward model says what the residual should look like. An agent staring at a deconvolved image adds hallucination risk without adding information the residual lacks. The obvious alternative is Morozov's discrepancy principle, used by Toader et al. for spatially varying light-sheet deconvolution. It matches only the total size of the residual to the expected noise level. It is blind to structure in the residual, such as edges of cells left behind by oversmoothing or ringing from a wrong blur kernel, and it gives one global number rather than a map.
Whiteness-based parameter choice is known in classical regularization, but nothing in the file applies it locally, for mixed Poisson–Gaussian noise, as a failure detector attached to learned models. That combination is what is new here.
Core Mechanism
Let $y$ be the raw image and $\hat{x}_\lambda$ the reconstruction at smoothing strength $\lambda$. Let $\mathcal{A}$ be the forward model, meaning the blur operator, possibly spatially varying as in Toader et al.. Let $a$ be the camera gain and $\sigma^2$ the read-noise variance. The normalized residual at pixel $i$ is
$$r_i(\lambda) = \frac{y_i - (\mathcal{A}\hat{x}_\lambda)_i}{\sqrt{a\,(\mathcal{A}\hat{x}_\lambda)_i + \sigma^2}},$$
which under a correct model is approximately zero-mean, unit-variance, and uncorrelated across pixels. For each patch $P$, let $\rho_P(k)$ be the sample autocorrelation of $r$ at spatial lag $k$, and $s_P^2$ its sample variance. We score the patch by
$$W_P(\lambda) = \sum_{0<|k|\le K} \rho_P(k)^2 + \big(s_P^2 - 1\big)^2,$$
where $K$ is a small lag radius, around 3 pixels. The global parameter is $\lambda^\star = \arg\min_\lambda \sum_P W_P(\lambda)$. The map of $W_P(\lambda^\star)$, calibrated against its null distribution from simulated pure noise, is the trust map.
The noise parameters $a$ and $\sigma^2$ are estimated once per acquisition from the data or from a calibration frame. The certificate then needs no ground truth. For denoising, $\mathcal{A}$ is the identity, so the test applies directly to the output of Noise2Noise or any other network.
Research Program
We will first implement the test around two variational solvers. The first uses Hessian Schatten-norm regularization, from Lefkimmiatis et al.: it penalizes image curvature, which avoids the blocky "staircase" artifacts of total variation, and is solved with primal–dual iterations. The second uses a plain $\ell_2$ (Tikhonov) penalty. For $\ell_2$ and nearby $\ell_p$ penalties, del Aguila Pla et al. prove the reconstruction changes continuously with the data. That matters here: a small change in $\lambda$ or in $y$ cannot make the certificate jump arbitrarily.
We will then study when the test fails. Its null distribution is exact only if the variance normalization is correct, so we must measure how errors in $a$ and $\sigma^2$ bias $\lambda^\star$. Heavy regularization also correlates residuals by construction, so we will quantify this effect and correct the null by simulation.
Next we will apply the certificate unchanged to learned denoisers and to a blurred-input network. The aim is to show it detects the failures that McCann et al. warn about in learned reconstructions under distribution drift.
Finally, we will package the solver, noise estimator, and certificate as a Model-Zoo-style description, extending the bioimage.io format with a forward-model field. We will also build an OME-Zarr task for Fractal so the trust map is computed alongside every run. This packaging is engineering work, not the scientific claim.
Baselines
The naive reference point is the raw image, together with the solver at a fixed default $\lambda$ chosen once on held-out data. Both are what a user without guidance gets. The upper reference is the oracle $\lambda$ chosen by peak signal-to-noise ratio (PSNR) against ground truth. The key metric is the gap to this oracle.
The strongest parameter-selection competitor is the discrepancy principle as used by Toader et al.. We will run it on the same solver, with the same noise estimates and the same $\lambda$ grid, so that only the selection rule differs. Beating it means a smaller mean oracle gap across all noise levels, and a clear win specifically under model mismatch, where we expect matching the residual norm to over-regularize.
The strongest reconstruction competitor is Noise2Noise on the FMD dataset. Zhang et al. report a mixed-test-set PSNR of 35.40 dB for it, against 34.88 dB for DnCNN and 32.71 dB for VST+BM3D. We do not expect a variational solver to beat 35.40 dB in-distribution, and our claim does not require it. For the failure-detection question, the fair competitor is Monte Carlo dropout uncertainty, the reliability map InstantDL offers non-programmers. We will train the same U-Net with dropout on the same data. Beating it means higher detection area-under-curve (AUC) for corrupted regions, and a higher correlation between the map and local error.
Two ablations isolate the key ingredient. The first drops the autocorrelation term, leaving a local variance-only test, which is essentially a patchwise discrepancy principle. The second drops the Poisson-dependent normalization and uses plain Gaussian residuals. If the full test does not beat the variance-only version, the whiteness idea adds nothing and the claim fails.
Why not just…?
Why not just use Noise2Noise or DnCNN, which are faster and score higher? On in-distribution FMD data they will likely win on PSNR, per the 35.40 dB figure from Zhang et al., and we say so. The certificate is not a competitor to them. It is a check we will attach to them. The experiment that settles its value is the shift test described below: if network outputs under shift carry residual structure that the certificate catches and dropout uncertainty misses, users gain something even when they keep the network.
Why not just use the discrepancy principle or SURE (Stein's unbiased risk estimate), which are already automatic? This is the real threat. If the discrepancy principle's oracle gap is as small as ours in every condition, the parameter-selection half of the proposal collapses to a known method. Only the trust map would remain as a contribution. SURE's Poisson–Gaussian variants were not in the file, so we cannot cite results. We will include one as an extra baseline if an implementation is available. The ablation against a local variance-only test is designed to make this outcome visible rather than hide it.
Why not just put an LLM agent in charge of tuning, as CellAgent and AutoMat do for their domains? AutoMat's reported gain, a formation-energy mean absolute error near 321 meV/atom versus over 2500 meV/atom for GPT-4.1 Vision, comes from a closed-loop physical check, not from the language model's visual judgment. AutoMat also attributes 39% of its failures to retrieving the wrong template. Our proposal supplies the physical check directly. As a small side experiment, we will give a multimodal LLM the reconstructions over the same $\lambda$ grid and ask it to choose. We will report its oracle gap alongside ours, and report it honestly if it wins.
Why not just wrap Richardson–Lucy or an existing solver in a GUI? A GUI moves the smoothing slider onto the screen, but the user still has no way to know where it should sit. Unregularized Richardson–Lucy also amplifies shot noise as iterations continue. The certificate can be dropped into such a GUI. The question it answers, whether this output is consistent with the physics, is one a slider cannot.
Proposed Experiments
The first experiment tests parameter selection. We use the FMD dataset from Zhang et al., which contains confocal, two-photon, and widefield images with 50-frame-average ground truth at noise levels $S \in \{1,2,4,8,16\}$. We blur the ground-truth images with known widefield and confocal blur kernels, then add Poisson–Gaussian noise matched to each level. We also run the pure-denoising case on the real noisy frames. We report the oracle PSNR gap for the default $\lambda$, the discrepancy principle, the variance-only ablation, and the full test.
As a sanity anchor, the Hessian Schatten-norm solver should reproduce its reported advantage over total variation. Lefkimmiatis et al. report ISNR gains above 0.5 dB, and 21.55 versus 18.52 dB on Boat in the 2%-pixels setting. We will check this on their benchmark before any certificate results are trusted.
The second experiment is the decisive check, detecting controlled mismatch. On the same data we introduce four corruptions: a blur kernel that is too wide or too narrow by 10–40%, a spatially varying aberration applied only to part of the field, a noise level shifted between training and test for the learned models, and a network applied to a microscope type it never saw, for example trained on confocal and tested on two-photon. For each, we compute detection AUC for flagging corrupted images, and the pixelwise correlation between the trust map and absolute error. We compare against Monte Carlo dropout and the variance-only ablation.
The proposal succeeds if the full certificate stays within roughly 0.3 dB of the oracle across noise levels. This tolerance is our target, not a number from the file. It must also clearly outperform dropout uncertainty on mismatch AUC. If both the discrepancy principle and dropout match it, we will report that and drop the claim.
The third experiment covers real light-sheet data. If the microsphere and Marchantia data from Toader et al. are obtainable, microspheres give a physical ground truth: bead width after deconvolution. We then ask whether a constant-blur model yields a trust map that lights up exactly where their spatially varying model says the blur changes. No numeric result is reported for that data (no reported result found on Lacuna for LS-IC SSIM values), so the comparison will be internal. A short study with a handful of facility biologists, asking whether they correctly reject deliberately corrupted reconstructions with and without the trust map, will test whether the certificate is actually readable.
References
- BioImage Model Zoo: A Community-Driven Resource for Accessible Deep Learning in BioImage Analysis
- CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis
- AutoMat: Enabling Automated Crystal Structure Reconstruction from Microscopy via Agentic Tool Use
- Image Reconstruction in Light-Sheet Microscopy: Spatially Varying Deconvolution and Mixed Noise
- 3D Poisson microscopy deconvolution with Hessian Schatten-norm regularization
- Stability of Image-Reconstruction Algorithms
- Biomedical Image Reconstruction: From the Foundations to Deep Neural Networks
- Fractal: Towards FAIR bioimage analysis at scale with OME-Zarr-native workflows
- A Poisson-Gaussian Denoising Dataset With Real Fluorescence Microscopy Images.
- InstantDL: an easy-to-use deep learning pipeline for image segmentation and classification