# PDR 004: Publish inspectable spatial evidence with aggregate scores

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/004-inspectable-spatial-evidence
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-06

## Context

An aggregate Dice score, F1 score or patient-level AUROC does not show whether a model found plausible morphology,
relied on preparation artifacts, missed small lesions, or failed on one scanner. Whole-slide outputs are spatial,
but a table can detach the score from the regions that produced it.

The repository already serves CPTAC-COAD pyramid tiles from R2 through a Leaflet viewer. Pathologize can extend
that working surface instead of starting with another general annotation or whole-slide viewing product.

## Proposed decision

Publish inspectable spatial evidence alongside aggregate benchmark results. A reader should be able to trace a
score to eligible slides, sampled regions, tissue and artifact masks, cell instances, model attention or region
scores, errors and exclusions where licences permit publication.

Use the existing Leaflet and R2 tile path as the first evidence viewer. Extend it with versioned experiment
overlays, run comparison and provenance when those capabilities are needed. Use interoperable spatial formats so
the same outputs can still be inspected with established pathology tools. Keep model attention labelled as model
behaviour, not as a causal pathological explanation.

Include representative failures and performance slices by dataset, scanner, stain, tissue, class and quality.
Do not publish patient material or derived images when the source licence does not permit it.

## Consequences

Researchers and potential contributors can inspect what changed rather than trusting one number. Spatial evidence
also makes tiling and quality experiments easier to compare.

Publication becomes heavier and may need restricted local inspection for some datasets. The current public viewer
also needs a deliberate boundary between redistributable examples and private or restricted artifacts. Large
overlays require versioned storage and efficient formats. Visual examples can encourage cherry-picking, so the
selection rule and complete machine-readable scorecard must accompany them.

## Evaluation

Accept the proposal when a baseline report lets a reader move from a summary metric to predetermined success and
failure cases, inspect their versioned spatial outputs, and recover the exact run that produced them without
exposing data outside its licence.

## Related records

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

---

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