Physically Anchored Plasmolysis Scoring for Measuring Osmotic Adjustment Across Labs

Plant cells under drought or salt stress protect themselves by packing more dissolved particles into their sap, so that water stays inside. This defense is called osmotic adjustment. It is almost always measured on ground-up tissue: an osmometer or a proline assay reports one number per leaf, which averages over every cell type and compartment. A much older microscope test can resolve individual cells. Tissue is bathed in a graded series of sugar-alcohol solutions, and the observer records at which concentration each cell's protoplast (the living contents inside the rigid wall) starts to pull away from the wall. This event is called incipient plasmolysis. It happens when the bath's osmotic strength matches the cell's own, so it reads out each cell's solute content on a physical scale. The weak point of the test is the person reading it. "Has this protoplast begun to retract?" is a borderline visual call on a continuous change, the same kind of call that produces moderate agreement among pathologists grading tissue proportions (Geessink et al. report κ = 0.578 for visual stroma scoring).

This proposal rests on one observation. Unlike most diagnostic imaging tasks, the plasmolysis series has a label that no observer controls: the osmolality of the bath, which is set by the experimenter and checked with an osmometer. We propose to train and evaluate cell-level plasmolysis scorers against that physical anchor instead of against human grades. We then ask whether anchored scorers keep their accuracy and calibration when frozen and moved to other labs, microscopes and species, where scorers trained on human grades are expected to fail. The result would be a spatially resolved, cell-by-cell map of osmotic adjustment. It would also be a clean test of a question computational pathology cannot run on its own data: how much of the "external validation drop" comes from the images changing, and how much from the local graders' threshold changing?

Core Claim

We make two claims. The first is that per-cell solute potential can be estimated from plasmolysis images with error small enough to resolve osmotic adjustment between treatments, and to resolve heterogeneity between cell types that bulk osmometry hides.

The second is that scorers supervised by bath osmolality transport across sites better than scorers of identical architecture supervised by local human consensus. The reason would be that human labels carry a site-specific decision threshold, and that threshold travels with the model. We expect this failure to appear mainly as poor calibration: the whole plasmolysis curve shifts to a different bath concentration. Ranking accuracy should suffer much less, which means area-under-curve (AUC) summaries alone would miss the problem.

The physics that makes the anchor possible is the van't Hoff relation at incipient plasmolysis:

$$\psi_s^{cell} \approx -\,c_{50} R T$$

Here $\psi_s^{cell}$ is the cell's solute potential, $c_{50}$ is the bath osmolality at which a cell has a 50% probability of being plasmolyzed, $R$ is the gas constant and $T$ is absolute temperature. Osmotic adjustment is then the drop in $\psi_s^{cell}$ between stressed and control cells of the same type.

The contribution is not a better neural network. If a simple geometric measurement does the job, that measurement becomes the method. What is new is the anchored supervision and the grader-by-site validation design.

Relation to Prior Work

Three lines of prior work bear on the proposal. The pathology work that joins observer variability and external validation supplies the evaluation discipline but also shows its blind spot. Shamai et al. froze a model that predicts breast-cancer receptor status from routine H&E slides and validated it across six cohorts and two scanner vendors, reaching external ER AUC 0.95. Yet the model's labels are themselves human reads of a second stain, and the model fails exactly in the 1–10% "ER-low" zone where readers disagree. Geessink et al. and Yang et al. replace visual tumor-stroma estimates with continuous automated scores that predict survival. However, Geessink still relies on human-chosen hot-spots and has no external cohort. In all three papers, ground truth is either human or a distant clinical outcome. Our setting removes the human from the label entirely, and we borrow their frozen-model, multi-site protocol.

The only work in the file that touches plant physiology and observer variability is Giuffrida et al.. It quantifies expert, non-expert and crowd disagreement on Arabidopsis leaf counts in wild type versus starch-deficient mutants. It finds that the average of three citizen raters matches an expert, while an automated counter falls outside the human error range on hard images. This paper motivates us and constrains us. It warns that a crowd consensus may already be good enough, so we test that head-on. But leaf counting has no physical ground truth, and the study used one imaging setup, so it cannot separate observer bias from domain shift.

The osmotic-adjustment literature supplies biological positive controls with known direction. Exogenous myo-inositol raises soluble sugars and proline in drought-stressed maize seedlings, so it should lower $\psi_s^{cell}$. Halogeton glomeratus raises leaf Na⁺ 3.18-fold under 200 mM NaCl, but bulk measurement cannot say where the ion sits. The cotton GhVacInv2a work shows vacuolar invertase driving turgor in one cell type during one developmental window, a signal that bulk assays dilute. These papers motivate cell-resolved measurement. They also give us treatments whose effects we can predict before looking at the images.

Research Program

We will build a multi-site plasmolysis image corpus. Epidermal peels or leaf sections come from control plants and stressed plants: salt, withheld water, and drought with and without myo-inositol pretreatment. Each sample is imaged live in a graded mannitol or sorbitol series of 10 to 14 concentrations, every concentration verified by freezing-point osmometry. Neutral-red vacuole staining is used where it helps.

Each participating lab uses its own microscope, camera, stain batch and growers. This deliberately reproduces the instrument and protocol shifts that break pathology models. Every sample also gets bulk sap osmometry and a proline assay.

For each cell we record three things. The first is the bath osmolality, which is the physical anchor. The second is a binary plasmolyzed/not call from at least three graders at the originating lab and three at a second lab, plus a crowd consensus collected in the Giuffrida style. The third is a segmentation of wall and protoplast.

The scorers output, for each segmented cell, a probability of being plasmolyzed at a given bath concentration. The anchored scorer fits each cell's probability curve under a monotone logistic model in bath osmolality, sharing parameters across the series. Its training signal is only the known bath value and the constraint that plasmolysis cannot reverse as the bath strengthens. No human call is used. The scorer's estimate of $c_{50}$ gives the cell's $\psi_s^{cell}$.

The human-supervised twin has the same backbone and is trained on local grader consensus. A classical scorer computes the protoplast-to-wall area ratio and wall-contact loss from the segmentation, then applies a threshold. All scorers are frozen at the derivation site before any external image is scored, following the protocol summarized in the external validation literature.

Baselines

The naive reference point is current practice. That is bulk sap osmometry for osmotic adjustment, and a single local expert reading the plasmolysis series by eye for per-cell $c_{50}$. It is the standard every alternative must beat. We will report its per-sample agreement with osmometry and its between-grader spread in $c_{50}$, measured in MPa.

The strongest human competitor is consensus scoring by three or more raters. Giuffrida et al. found that three non-experts matched an expert on leaf counts, and they report internal consistency α > 0.98. Consensus will be run on exactly the same cells, with raters blind to treatment and bath value.

The strongest machine competitor is the human-supervised deep scorer. It mirrors the multiple-instance deep pipeline that Shamai et al. validated externally. It gets the same backbone, training images, tuning budget and frozen threshold as the anchored scorer. It is also the key ablation, because the only difference between the two is the source of the label. No reported result exists on Lacuna for any plasmolysis-scoring task. The nearest reported numbers come from unrelated tasks: ER AUC 0.95 from Shamai et al., and 96.47% VGG16 versus 86.08% random-forest accuracy on the gastric patches of GasHisSDB. These indicate what deep versus classical gaps look like elsewhere, not what to expect here.

The classical geometric scorer is the cheap competitor. It gets the same segmentation masks and a threshold tuned on derivation data only.

We count the anchored scorer as beating a baseline under two conditions. First, its external-site error in $\psi_s^{cell}$ against the bath-derived truth must be lower, with a bootstrap confidence interval over plants excluding zero. Second, its calibration slope must stay closer to 1. We fix in advance that an equal AUC does not count as a win or a loss.

Why not just…?

Why not just measure osmotic adjustment with an osmometer? For whole-tissue averages it is the right tool, and we use it as our coarse check. It cannot say whether guard cells, pavement cells or bundle-sheath cells adjusted, nor whether the extra Na⁺ in Halogeton sits in vacuoles. This is a real threat if cell-to-cell heterogeneity turns out to be small. So we explicitly estimate the between-cell-type variance in $\psi_s^{cell}$. If the cell-resolved map adds no variance beyond bulk osmometry plus measurement noise, we will report that the imaging route is unnecessary for these treatments.

Why not just average three human raters? Giuffrida et al. make this a serious option. Our design tests it directly: consensus calls are compared against the physical anchor at borderline bath concentrations and across sites. If consensus $c_{50}$ matches bath-derived $c_{50}$ equally well at every lab, the case for anchored models reduces to cost and throughput. We expect consensus to still carry a lab-specific threshold, because all three raters are trained in the same local convention. The grader-crossing experiment below measures exactly that.

Why not just threshold the protoplast-to-wall area ratio? Plasmolysis is largely geometric, so this may work as well as a deep model, and we treat it as a genuine candidate rather than a straw man. It is also fully compatible with anchoring: its threshold can be fit to bath osmolality instead of human calls. If it matches the anchored deep scorer externally, it becomes the recommended method. The claim about label source would still stand, because it concerns where the threshold comes from, not which estimator implements it.

Why not just normalize stain color and skip multi-site collection? Color normalization addresses only one kind of shift. It cannot tell us whether graders at different sites draw the line at different retraction depths. Only multiple sites, crossed with multiple grader groups, can separate those two effects.

Proposed Experiments and Decisive Checks

The first check is biological validity. Anchored $c_{50}$ must detect the expected direction of osmotic adjustment in the myo-inositol and salt treatments. Its per-plant average must also track bulk sap osmometry. Failing either check invalidates the anchor, whatever the image metrics say.

The decisive experiment crosses images and graders. Images from site A are graded by graders from A and from B, and likewise for site B. We fit a model of each grader group's implied $c_{50}$ with terms for image site and grader site. If the grader-site term is comparable to or larger than the image-site term, much of what pathology calls external-cohort failure would, in this analogous setting, be label shift. The human-supervised scorer's external calibration error should then track that term, and the anchored scorer's should not.

Transport is tested on held-out labs and a held-out species, with all models frozen in advance. We report error in $\psi_s^{cell}$ (MPa), calibration slope and intercept, and Brier score, with AUC reported only as a secondary metric. A final ablation feeds the anchored scorer a label source that is increasingly corrupted by one site's human calls. This asks how quickly external calibration degrades as human threshold bias enters supervision, which bears directly on how pathology models trained on subjective labels should be expected to travel.