When a tissue-analysis model trained on human slides is applied to dog or rat tissue, researchers usually report an average score on a small labeled sample and call the transfer "good enough." The first problem is that the average hides how often the model gets a slide badly wrong. The second is that the score mixes two different causes. Some failures come from biology: the tissue of another species really looks different. Others come from the scanner, since the same slide digitized on different hardware already breaks segmentation models. We propose to borrow the tool particle physicists use when a search finds nothing. In a rare-decay search, the experiment counts candidate events, measures how many would appear from background noise alone in a separate control sample, and reports a 90% confidence ceiling on the true signal rate, called a branching fraction upper limit. Here the "signal" is failures caused by the species gap, the "background" is failures caused by scanner hardware alone, and the output is a statistically valid ceiling on how often a model fails because of species.
The payoff is a decision rule rather than a descriptive score. If the ceiling on species-caused failure falls below a tolerance the downstream study can live with, the model can be used in the new species without retraining. If the limit sits above the tolerance while the scanner background is large, the right fix is scanner normalization, not species-specific relabeling. A second ingredient comes from temporal query decomposition. Longitudinal preclinical conclusions of the form "collagen rose between day 7 and day 14 and treatment slowed it" can be split into per-timepoint measurements plus a temporal operator. The ceiling can then be set on the abundant per-timepoint measurements and carried through to the rare compound claim, instead of being estimated directly from a handful of time-course comparisons.
Core Claim
A counting model with a separately measured background, which physics calls the "on/off" problem, gives valid and useful upper limits on the species-attributable failure rate of a histopathology model. It can use very few labeled target-species slides and still does not confuse scanner artifacts with biology. When applied per timepoint and propagated through temporal operators, it gives tighter valid bounds on longitudinal conclusions than bounding those conclusions directly.
A tissue region counts as a failure when the model's morphometric readout deviates from the expert annotation by more than a tolerance $\tau$. Examples of readouts are the tumor area fraction or the collagen area percentage. In the target species on a given scanner, we observe $n$ failures among $N$ regions. In a control sample of $M$ region pairs, we observe $m$ cases where the same physical tissue, scanned on two different machines, changes its readout by more than $\tau$. That control sample measures the scanner-only failure rate $b$. The model is
$$ n \sim \mathrm{Binomial}(N,\ s + b), \qquad m \sim \mathrm{Binomial}(M,\ b), $$
where $s \ge 0$ is the species-attributable failure rate. The quantity we report is the 90% confidence-level upper limit on $s$, obtained with $b$ profiled out. This is the same construction that turned a count of zero or few candidates into ceilings such as $\mathcal{B}(\tau\to\mu\gamma) < 1.0\times10^{-6}$ in the CLEO search and the $10^{-5}$–$10^{-6}$ limits of BABAR's hadronic-tag B searches. There, exclusive tagging and a background fit replaced naive counting. We reuse the statistics, not the physics. The additive split assumes that scanner and species failures do not overlap much. We test that assumption directly, as described below.
Research Program
Prior work plays three roles in this plan.
Work that motivates the proposal. Wilm et al.'s multi-scanner canine SCC dataset digitizes 44 canine squamous cell carcinoma samples on five scanners. It shows in-domain segmentation mIoU of 0.82–0.86 but cross-scanner drops of up to 0.38. That paper treats scanner shift purely as a computational benchmark. We use its repeated scans of the same tissue as the control region that measures $b$, which that paper never does. Cross-species studies such as the CMLPS-N1 canine tumor model argue that spontaneous canine disease is a better human proxy. They report positive findings but give no quantitative ceiling on mismatch, which is the gap this proposal fills. Rat fibrosis studies using unilateral ureteral obstruction and bile duct ligation (Zhong et al.; methane-rich saline in cholestasis) score collagen at single endpoints and acknowledge the translation gap without quantifying it.
Work that supplies the method. Physics supplies the upper-limit construction. Temporal decomposition supplies the propagation step. Systems such as TEQUILA and Saquete et al. split a question into atemporal sub-questions that are answered first and a temporal operator applied afterward. Lutz et al.'s propositional abstraction checks temporal formulas over boolean atoms that are evaluated separately. We apply the same idea to a longitudinal claim. Each atom is "the readout at timepoint $t$ in animal $a$ is within $\tau$." The claim is a temporal formula over those atoms. A per-atom ceiling $\epsilon$ bounds the claim's failure probability by at most $k\epsilon$ for a claim that depends on $k$ atoms, by a union bound.
Work that constrains the design. The LTLZinc neuro-symbolic study found two things. Neural-only sequence models failed to learn temporal labels, and post-hoc temperature calibration had limited effect compared with architecture choices. We therefore do not try to calibrate a monolithic time-course model. That study also warns that errors at the perception stage cascade downstream. This is exactly what our per-atom limits quantify.
What we build. First, a source model: a tissue segmenter trained on human data from OCELOT, which spans six human organs and 306 slides and which we use for cancer-area segmentation. Second, the control-region estimator for $b$, built from same-tissue scanner pairs in the canine dataset. Third, the profile-likelihood limit on $s$, stratified by target scanner. Fourth, the propagation layer for temporal claims.
The temporal study needs longitudinal data that the file does not provide. We would collect a small rat unilateral-ureteral-obstruction time course, stained with Masson's trichrome at three timepoints with treated and control arms. Each slide would be scanned on two machines so that $b$ can be measured within the rat data. This is new data collection, and we flag it as the costliest part of the plan.
Baselines
The naive reference point is the practice we want to replace. It reports mean mIoU of the human-trained model on canine tissue and treats a drop from the in-domain 0.82–0.86 as the "species gap." It runs on the same canine slides with no extra machinery. We beat it if our method makes decisions it cannot make, most importantly correctly attributing a large failure count to scanner rather than species on scanners like the Pannoramic 1000. Wilm et al. report that this scanner's white-background thresholding causes cross-scanner failure.
The strongest statistical competitor is an exact binomial upper bound (Clopper–Pearson) on the raw failure fraction $n/N$, with no background term. It uses the same labeled regions, the same $\tau$, and no tuning, so the comparison is exact. It is a valid ceiling on total failure. We beat it only in two cases. The first is when our limit on $s$ is substantially lower where the scanner background is large, while still covering the true $s$ at 90% in injection tests. The second is when it leads to a different and correct action, meaning "normalize" instead of "relabel."
For the temporal claims, the direct baseline applies the same exact bound to observed failures of whole longitudinal claims. The decomposition wins if per-atom propagation gives a tighter bound at equal coverage.
The strongest model-side reference, which fixes what "good" means in-domain, is the best reported numbers. These are the 0.86 in-domain canine mIoU from Wilm et al. and the OCELOT tissue-context model's mean F1 of 71.23, up from 64.44, in OCELOT. There is no reported result on Lacuna for human-to-canine transfer of a tissue segmenter, so that number will be produced here.
Two ablations isolate the key ingredients. The first removes the control region by setting $b=0$, which reduces the method to the raw exact bound. The second deliberately mis-estimates $b$ using a different scanner pair, which tests how sensitive the method is to background mismodeling. Physics identifies background mismodeling as the dominant systematic at high statistics.
Why not just…?
Why not just report a bootstrap confidence interval on target-species mIoU? A bootstrap bounds the average overlap, not the rate of slides that fail beyond a tolerance, and a downstream morphometric study is damaged by the tail. It also has no term for scanner background, so it inherits the attribution problem of the naive baseline. We include bootstrap intervals in the experiments and show how often they approve transfers where more than $\tau$-level failures occur in over 10% of regions.
Why not just use the raw exact binomial bound? This is a real threat. If a user only needs to know "how often will this model fail on my canine slides with my scanner," the raw bound answers that correctly and more simply, and we will say so. Our method earns its place only when the question is why the model fails, which decides between retraining and normalization. It also matters when the target lab's scanner differs from the validation scanner, so that only the species term transfers. The scanner-stratified experiment settles this. If the limits on $s$ do not change the correct action in any stratum, the extra machinery is not justified.
Why not just use conformal prediction on a labeled target calibration set? Conformal methods give per-prediction coverage under exchangeability between calibration and test data. They need labeled target data just as we do, and they do not separate causes of error. They are complementary rather than competing. Still, a reviewer could argue that conformal intervals on the readout already flag bad regions. We will run split-conformal on the same calibration regions and compare the fraction of flagged regions against our ceiling, and we will report whether conformal flags track scanner rather than species.
Why not just train an end-to-end model on the time course? The LTLZinc results ( A Neuro-Symbolic Framework for Sequence Classification with Relational and Temporal Knowledge) show neural-only temporal models failing to learn sequence labels. More basically, a monolithic model gives no per-timepoint atoms to bound, so no ceiling on the compound claim is available at all.
Proposed Experiments
The first experiment is calibration by signal injection, the check physics uses to validate a limit-setting procedure. On canine tissue scanned with the reference scanner, we inject a known species-like failure rate $s^\ast$ by applying controlled morphology perturbations to a random subset of regions. We then verify that the 90% upper limit on $s$ covers $s^\ast$ in at least 90% of repetitions across all five scanners. We also report the limit's median width against the raw exact bound. The overlap assumption behind the additive model is checked here by injecting failures into regions that already fail from scanner shift. If coverage breaks there, we replace the additive rate with a noisy-OR combination.
The second experiment is real human-to-canine transfer. We apply the OCELOT-trained segmenter to canine SCC tumor regions and report per-scanner $n$, $m$, the limit on $s$, and the raw bound. We also check whether the limit on $s$ stays stable across scanners, as a species quantity should, while raw failure rates vary by the up-to-0.38 shift already documented. A limit that stays stable across scanners is the decisive evidence that the background model separated the two causes. If it swings with the scanner, the method has failed.
The third experiment is the temporal propagation study on the new rat time course. We compare propagated bounds on claims like "treated collagen area is below control at day 14 and rose less between day 7 and day 14" against direct bounds on those claims. We verify coverage by holding out annotations. The expected result is that propagation is tighter when atoms far outnumber claims. If the union bound's factor $k$ makes it looser than the direct bound for realistic claim lengths, then the decomposition half of the proposal does not pay off and should be dropped in favor of the single-timepoint limit alone.
References
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Searches for the decays<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>l</mml:mi><mml:mo>±</mml:mo></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>∓</mml:mo></mml:msup></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>l</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>ν</mml:mi></mml:math>(<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>l</mml:mi><mml:mo>=</mml:mo><mml:mi>e</mml:mi></mml:math>,<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>μ</mml:mi></mml:math>) using hadronic tag reconstruction
- Multi-scanner Canine Cutaneous Squamous Cell Carcinoma Histopathology Dataset
- CMLPS-N1: a novel preclinical cell line model for canine mammary tumor and its application in therapeutic screening
- All-trans retinoic acid regulating angiopoietins-1 and alleviating extracellular matrix accumulation in interstitial fibrosis rats
- Methane-Rich Saline Counteracts Cholestasis-Induced Liver Damage via Regulating the TLR4/NF-κB/NLRP3 Inflammasome Pathway
- TEQUILA: Temporal Question Answering over Knowledge Bases
- Splitting Complex Temporal Questions for Question Answering Systems
- LTL over Description Logic Axioms
- A Neuro-Symbolic Framework for Sequence Classification with Relational and Temporal Knowledge
- OCELOT: Overlapped Cell on Tissue Dataset for Histopathology
- A Neuro-Symbolic Framework for Sequence Classification with Relational and Temporal Knowledge