Tissue engineers often build implants from animal or donor tissue. They wash out the cells, which carry the immune-triggering "xenoprotein load", and keep the collagen-and-laminin skeleton (the extracellular matrix, or ECM). Whether a scaffold is clean enough, and whether the washing damaged the matrix, is judged mostly by eye on stained tissue sections plus a bulk DNA assay. We propose to train an image model that reads a standard hematoxylin-and-eosin (H&E) section of a decellularized scaffold and scores it on two separate axes. The first is residual cellular material. The second is matrix damage. Because no lab produces more than a handful of such slides, the training data must be assembled from figures scattered across hundreds of papers, which makes this a small-data problem much like rare-disease pathology.
The second, methodological claim is about evaluation. When a dataset is stitched together from the literature, the same lab's staining, scanner and protocol recur across its papers. A model can learn to recognize the lab and pass for a model that recognizes the tissue. We therefore hold out whole labs, identifying them by disambiguating author names across papers. We also argue that the usual error trade-off in author disambiguation runs the wrong way for this purpose. The same evaluation discipline then carries over to rare-disease case-report corpora, where one group often publishes many cases.
Core Claim
Our first claim is that a pathology foundation model, meaning a large image encoder pretrained without labels on millions of slides, can be adapted with very few labels to score decellularization quality from H&E. Its advantage over simply counting nuclei should appear on the axis that nuclei counts cannot see: matrix damage from harsh detergents. The second claim is that honest evaluation on literature-mined pathology data requires splits at the level of the lab, not the paper. The tool that builds those splits, author disambiguation, should be tuned to avoid splitting one lab into two clusters, because a split lab leaks images across the train/test boundary. Merging two distinct labs into one cluster only discards some usable data.
Why This Gap Is Real
Every decellularization study in the file uses histology as qualitative confirmation. The thyroid acellular scaffold study verified clearance with H&E, Masson's trichrome (a stain that turns collagen blue) and electron microscopy, alongside a residual DNA of 49 ± 23 ng/mg. The decellularized epineurium study graded clearance and laminin preservation manually. The high-speed centrifugation osteochondral work confirmed marrow and lipid clearance histologically. The denuded amniotic membrane study counted macrophages by hand. These papers motivate the proposal: they show that histology is the accepted evidence, that it is read subjectively, and that the field's central trade-off is between thorough clearance and matrix destruction. Stronger SDS or longer trypsin removes more DNA but strips glycosaminoglycans and disorders collagen. No single DNA number captures that trade-off.
On the method side we reuse rather than invent. Virchow is a 632M-parameter encoder self-supervised on 1.5M H&E slides. It reports a specimen-level AUC of 0.937 on seven rare cancers, which is evidence that general morphology features transfer to classes with little data. KEEP aligns an image encoder with a disease knowledge graph and supports zero-shot prediction from text. It reports 0.898 sensitivity at 0.95 specificity for zero-shot cancer detection and an 8.5-point balanced-accuracy gain over CONCH on rare brain tumor subtyping. Neither model was evaluated on engineered or acellular tissue, and that is the domain shift we test.
LoGex synthesizes rare-class tiles with a guided diffusion model. On its long-tail benchmark it lowers FPR95, the false-positive rate at 95% true-positive rate, from 25.05% (cross-entropy) to 16.25%. It also constrains our design: its synthetic tiles can contain biologically implausible structure. For scaffolds, that is exactly the failure that matters, since a generator that hallucinates clean matrix would teach the model to approve bad scaffolds.
On the attribution side, the PubMed disambiguation system has four components. It scores pairs of name occurrences with a regularized classifier over coauthors, MeSH terms, venues and affiliations, and calibrates those scores with Pool Adjacent Violators. It then clusters name occurrences into author profiles, correcting transitivity violations along the way. It was tuned for search: lumping (wrongly merging two people) fell from 10.1% to 2.2%, while splitting (one person divided into several profiles) rose from 1.8% to 7.7%. That is the right choice for bibliometrics and the wrong one for leakage control, and our split construction re-tunes the same method toward the opposite error. The many author-correction notices in the file, such as the liver cell atlas correction and the PAFAH2 kidney injury correction, show that raw name strings are unreliable keys even within histopathology papers. Grouping by exact name string would therefore under-merge.
Research Program
We will build two datasets. The first is a literature-mined corpus of H&E panels from open-access decellularization papers. Each panel is paired with whatever the paper reports: native-versus-decellularized status, detergent regimen and exposure time, residual DNA where given, and the tissue of origin. The labels are noisy and weak by construction. The second is a small controlled "anchor" set that we produce ourselves. One tissue type, porcine or rat dermis or thyroid, is decellularized along a graded ladder of SDS concentration and exposure time, with Triton X-100 and sodium deoxycholate arms modeled on the epineurium protocol. Every sample is scanned as a whole slide and measured for DNA, glycosaminoglycans and collagen content. The anchor set is the only place where both axes have real ground truth. It is generated by one lab, so it also serves as a fully out-of-distribution test for anything trained on the corpus.
Each corpus image is assigned a lab identity. We take author lists and affiliations from the source papers and run a PubMed-style pairwise classifier with calibrated probabilities. We then cluster with a low merge threshold, so that ambiguous pairs are joined rather than separated. We report the resulting split under three groupings: random, by paper, and by lab. Comparing these isolates how much apparent accuracy is really lab recognition.
The model itself is deliberately simple. It consists of frozen foundation-model tile embeddings, a small attention pooler over tiles, and two output heads, one for residual cellular load and one for matrix damage. It is trained on the corpus's weak labels and calibrated on a few anchor slides. The novelty is in the two-axis target and the evaluation, not in the architecture.
Baselines
The naive reference is color deconvolution followed by nuclei counting: hematoxylin is separated computationally and nuclei per mm² are thresholded. A matching trichrome-free proxy for matrix damage is eosin intensity and texture statistics. This is what a careful lab would do with no machine learning. It is the strongest competitor on the clearance axis, because hematoxylin directly marks nucleic acid. It is run on the same slides, with its threshold tuned on the same calibration slides the learned models see.
The strongest learned baseline is a linear probe on Virchow embeddings. It uses the same tiles, the same labeled budget and the same hyperparameter search, which is 20 configurations for every method. A second learned baseline is KEEP used zero-shot with text prompts describing acellular matrix, residual nuclei and disrupted collagen, and then few-shot with the same label budget.
No result on decellularization scoring was found on Lacuna for any of these models. The closest reported figures are on rare cancers rather than on this task: Virchow's 0.937 AUC and KEEP's 0.898 sensitivity at 0.95 specificity. They indicate transfer potential, not a target to beat. We count the proposal as succeeding on clearance only if it matches the nuclei count within its confidence interval while also beating it on matrix damage. On matrix damage it must achieve a higher rank correlation with measured glycosaminoglycan loss and with detergent dose on the lab-held-out anchor set, by a margin that excludes zero under bootstrap. The method's own gain over the linear probe must also survive the lab-held-out split.
Two ablations isolate the key ingredients. The first trains with matrix-damage labels removed, which tests whether the second head adds anything. The second evaluates every method under random, paper-grouped and lab-grouped splits, which tests whether lab grouping changes conclusions. A LoGex-style augmentation arm is included as an alternative route to coping with tiny labeled classes.
Why not just…?
Why not just count nuclei? For residual DNA this may be as good as anything, and we expect it to be very hard to beat on the clearance head. What it cannot see is matrix damage, which is the half of the field's trade-off that DNA thresholds such as 50 ng/mg ignore. If the anchor-set correlation with glycosaminoglycan loss is no better for the foundation model than for eosin texture statistics, the first claim is false, and we would report a lab-grouped benchmark plus the nuclei-count tool as the contribution.
Why not just group by paper instead of by lab? If labs reuse staining and scanners, paper grouping leaks. We test this directly in two ways. We measure the accuracy drop from paper-grouped to lab-grouped splits, and we train a probe to predict lab identity from embeddings. If the drop is negligible and the lab probe is near chance, author disambiguation is unnecessary here and we will say so. That outcome is plausible for strongly stain-normalized data.
Why not just use exact author-name matching, or the existing disambiguation system as is? Name-string matching fails on inversions and misspellings, which the correction notices show are common. The off-the-shelf system's tuning toward splitting, at a 7.7% splitting rate, leaves residual leakage. We compare leakage under three settings: exact matching, the default operating point, and our merge-favoring threshold. Leakage is measured as the lab-probe accuracy across the train/test boundary.
Why not just synthesize more scaffold images with guided diffusion? LoGex shows that synthetic tiles help classifier boundaries, but also that they can contain implausible structure. For a quality-control model, hallucinated clean matrix is the dangerous error. The augmentation arm shows whether synthesis helps on lab-held-out anchor slides or merely on in-distribution corpus slides.
Proposed Experiments
The main experiment trains every method on the literature corpus and evaluates in two places. One is lab-held-out corpus folds, scored with AUROC for native-versus-decellularized and for the reported DNA pass/fail when available. The other is the anchor ladder, scored with Spearman correlation of each head against measured DNA, glycosaminoglycans and detergent dose.
The decisive check has three parts. The foundation model must beat nuclei counting on matrix damage on the anchor set. Its ranking against the Virchow linear probe must be the same under lab-grouped and random splits. The lab-identity probe must lose accuracy after lab-grouped training, showing the model is not simply recognizing labs.
A transfer experiment applies the identical split protocol to rare-disease case-report figures, such as juvenile xanthogranuloma and stratified mucin-producing cervical carcinoma panels drawn from the file's case-report literature. It fine-tunes KEEP few-shot and measures how much the reported accuracy shrinks when cases from the same author group are held out together. If that shrinkage is large, it suggests that published few-shot rare-disease results built from literature images are inflated by the same lab-recognition effect.
References
- Development of a biomimetic thyroid acellular scaffold as a 3D platform for modeling thyroid cancer aggressiveness and drug resistance
- Preparation and evaluation of decellularized epineurium as an anti-adhesive biofilm in peripheral nerve repair
- High‐Speed Centrifugation Efficiently Removes Immunogenic Elements in Osteochondral Allografts
- Tissue remodeling after ocular surface reconstruction with denuded amniotic membrane
- A foundation model for clinical-grade computational pathology and rare cancers detection
- Knowledge-enhanced pretraining for vision-language pathology foundation model on cancer diagnosis
- LoGex: Improved tail detection of extremely rare histopathology classes via guided diffusion
- Author name disambiguation for PubMed
- Author Correction: Towards a reference cell atlas of liver diversity over the human lifespan
- Author Correction: PAFAH2 suppresses synchronized ferroptosis to ameliorate acute kidney injury