Discovery
PLISM is stronger than the other public candidates for scanner-domain feasibility, but its preparation contrast has a stricter identifiability limitation than the current registry wording makes explicit.
The primary paper states that 13 separate slides were stained using the 13 selected H&E conditions. Because mutually exclusive staining conditions require serial sections, each preparation label is attached to a different physical section/slide. Therefore preparation condition is perfectly aliased with section identity in the observed dataset.
The same stained section is then imaged across multiple scanner domains, so scanner effects can be studied with same-section pairing. Preparation effects cannot be cleanly separated from serial-section morphology, sectioning damage, local tissue loss, or slide-specific handling.
Primary source: https://doi.org/10.1038/s41597-024-03122-5
Consequence
Any PLISM experiment should distinguish two estimands:
- Scanner estimand: within-section, paired across scanner domains. This is the strongest supported contrast.
- Preparation-plus-section estimand: across serial sections with different stain conditions. This is not an isolated preparation effect.
A representation model that predicts stain condition may be learning true staining variation, serial-section morphology, tissue loss, registration residuals, or combinations of these.
Required safeguards
- Print a deterministic aliasing warning whenever
stain_condition -> section_id is one-to-one.
- Never label the serial-section contrast as a causal or isolated preparation effect.
- Use same-section scanner contrasts as the primary positive control.
- For preparation analyses, aggregate over many registered coordinates and report sensitivity to registration error and tissue dropout.
- Treat section identity as the experimental unit, not individual tiles.
- Split and bootstrap by registered field / tissue core and section hierarchy rather than by tile.
- Consider a crossed mixed-effects or paired-difference analysis only after the manifest proves repeated tissue cores across all 13 sections.
Implementation target
Add this check to the future benchmarks/plism_feasibility/ manifest audit and revise the public-dataset discovery wording from a generic missing-provenance limitation to explicit structural aliasing.
Discovery
PLISM is stronger than the other public candidates for scanner-domain feasibility, but its preparation contrast has a stricter identifiability limitation than the current registry wording makes explicit.
The primary paper states that 13 separate slides were stained using the 13 selected H&E conditions. Because mutually exclusive staining conditions require serial sections, each preparation label is attached to a different physical section/slide. Therefore preparation condition is perfectly aliased with section identity in the observed dataset.
The same stained section is then imaged across multiple scanner domains, so scanner effects can be studied with same-section pairing. Preparation effects cannot be cleanly separated from serial-section morphology, sectioning damage, local tissue loss, or slide-specific handling.
Primary source: https://doi.org/10.1038/s41597-024-03122-5
Consequence
Any PLISM experiment should distinguish two estimands:
A representation model that predicts stain condition may be learning true staining variation, serial-section morphology, tissue loss, registration residuals, or combinations of these.
Required safeguards
stain_condition -> section_idis one-to-one.Implementation target
Add this check to the future
benchmarks/plism_feasibility/manifest audit and revise the public-dataset discovery wording from a generic missing-provenance limitation to explicit structural aliasing.