# PDR 005: Evaluate restoration as reconstruction and model input

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/005-restoration-as-reconstruction-and-model-input
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-06

## Context

Super-resolution can make a low-resolution pathology image look sharper while inserting, removing or changing
cellular structure. Pixel similarity alone may reward a visually convincing image that is unsafe as biological
evidence. Improvement in one downstream classifier can also reflect artifacts introduced for that model.

Resolution reconstruction remains a useful research question for compression, remote review, older scans and
multiresolution model efficiency. Learned restoration may also expose a stronger feature signal when downstream
models receive blurred tiles, corrupt or missing tiles, or unavailable pyramid resolutions.

## Proposed decision

Evaluate image restoration in two explicit roles: an uncertain reconstruction for inspection and a transformation
or feature extractor for downstream models. Keep the original image, degradation or missingness mask, and available
native resolutions immutable beside every derived artifact. Record generated pixels, latent features and
transformations in the artifact provenance.

Require paired or defensible synthetic reference data and measure structural insertion and deletion, nucleus and
cell consistency, morphometry, embedding drift, cross-scanner behaviour, uncertainty and downstream-task effects.
Compare learned restoration with the degraded input, conventional interpolation or inpainting, and models trained
directly for robustness to the same corruption. Do not promote a method from PSNR, SSIM or visual preference alone.

Reconstructed images carry no source-observation authority and cannot become source annotations or benchmark
ground truth. A restoration model may still be useful as an internal representation when sealed downstream
evaluation shows a robust signal gain. A later use outside research requires its own decision and evidence.

## Consequences

The project can study resolution, missing data and bandwidth trade-offs while preserving which morphology was
observed. The evaluation may reveal a restoration or latent representation that helps a downstream model despite
being unsuitable for human interpretation, or the reverse.

Paired pathology scans are difficult to obtain and register. Strong validation may cost more than training the
model, and some hypotheses will remain untestable with public data. The boundary deliberately rejects attractive
demonstrations that cannot measure hallucinated structure.

## Evaluation

Accept the proposal only after a benchmark detects known synthetic insertions and deletions and reports image,
morphology and downstream-task failures. It must include controlled blur, missing tiles and unavailable-resolution
cases, then test any apparent signal gain on sealed data with naturally occurring quality variation. Do not begin
a public model comparison until those guardrails exist.

## Related records

- Project: [Pathologize](https://robbiepalmer.me/projects/pathologize.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt