← Back to the vault

Lab Note · updated 2026-06-20

Image Quality Is Not One Number

Fovea treats microscopy quality as four nested levels — image-space, run/sample, readout, and decision — because a metric at one level says nothing about the levels above it. Full-reference scalars like SSIM and PSNR measure only the first, yet teams report as if they measured the third.

QCReportCell PaintingDigital PathologyLight-Sheet 3D/4DSpatial OmicsSSIMPSNRFIDno-reference IQAμDeepIQA
← Back to constellation

The problem — "Image quality" is routinely collapsed into a single full-reference scalar — SSIM or PSNR — computed against a clean ground truth. Two assumptions fail in quantitative microscopy. First, the reference rarely exists: the gentle, photon-limited acquisition is the only measurement, with no noise-free target to difference against. Second, even granted a reference, these metrics are intensity-weighted and spatially global — they quantify pixel similarity, not the estimand each platform reports: a morphological profile, a segmentation mask, a transcript count, a slide-level label. A pipeline can hold SSIM fixed while the quantity it exists to measure drifts.

The Fovea model — quality is four nested levels. Each level gates the next, and a number at one level says nothing about the levels above it.

  • Image-space quality — the per-image physics: resolution, SNR, contrast, focus, illumination uniformity, high-frequency content. This is where SSIM/PSNR live and where they are most useful but most reference-hungry. Do it better by decomposing into interpretable markers (Corbetta & Bocklitz), predicting quality per-patch with learned, out-of-distribution-robust IQA (μDeepIQA), or ranking focus blind with no-reference sharpness (Catanante et al.) — the same reference-free signal that drives measure where it matters.
  • Run / sample quality — consistency across the acquisition, not one frame: field-to-field drift, batch effects, z-depth signal decay, staining variation, registration stability. A perfectly sharp field can sit inside a batch-confounded run.
  • Readout quality — fidelity of the quantity the platform actually reports: segmentation accuracy, transcript assignment, profile reproducibility (percent-replicating / mAP), slide-level prediction, lineage tracking. This is the only level the science consumes.
  • Decision quality — the operational verdict the pipeline must emit: is this result reportable, does it need human review, or should it be rejected and re-acquired? QC exists to drive this level, and it is what a QC-aware report surfaces.

The levels are causal — image-space feeds run, run feeds readout, readout feeds decision — yet most pipelines instrument only the first and report as though they had measured the third.

Where it breaks — Because the levels are only loosely coupled, optimizing image-space quality does not propagate up to readout quality, and the decoupling is modality-specific — a denoiser, deconvolution, or super-resolution model can raise the score while corrupting the estimand. In Cell Painting profiling, oversmoothing lifts SSIM but erodes the texture and intensity statistics that encode phenotype, collapsing profile reproducibility. In digital pathology, focus, staining, and compression artifacts can leave global scores acceptable while shifting the feature distribution a foundation model embeds, degrading slide-level inference. In light-sheet volumes, destriping and deconvolution improve PSNR yet can synthesize structure that distorts segmentation and lineage tracking in deep, low-SNR regions. In spatial omics, quality propagates through segmentation and registration, where sub-pixel error misassigns transcripts and corrupts the count matrix. The methodological literature generalizes the warning: misaligned metrics and evaluation noise let benchmark gains outrun real utility (Varoquaux & Cheplygina; see Anderson et al. for a task-based audit that overturned standard IQA rankings). The quality that survives is task-conditioned — defined against the readout and the decision, not the pixels — which is why it belongs with verified outputs.

A higher SSIM or PSNR after denoising or super-resolution is not evidence the biology survived — it can be evidence the metric was optimized. Gate on the readout and the decision — profile reproducibility, segmentation/registration accuracy, classification — not on image-space similarity.

References

References

Local graph