Validating whole-slide quantification of hepatic vacuolation in aquatic models: correlational agreement is standing in for accuracy

Question: How are deep learning and whole slide image analysis pipelines being validated for automated quantification of hepatic vacuolation and other non-alcoholic steatohepatitis-like histopathology in fish and other non-mammalian aquatic model species?

Answer

Taken together, the retrieved literature shows that this question has almost no direct evidentiary base: exactly one retrieved pipeline quantifies hepatocellular vacuolation in a fish species on whole slide images, and it validates by showing that its continuous output is coherent with a pre-existing qualitative histological assessment on the same samples rather than by measuring segmentation accuracy, reader agreement, or repeatability [6]. Everything else that resembles validation machinery — nnU-Net segmentation with correlation to ordinal scores, foundation-model embeddings benchmarked against pathologists, multiple-instance-learning benchmarks, non-inferiority testing against reader consensus, pre-registered external validation — exists in human and murine MASH/MASLD work and has not been ported to aquatic or non-mammalian models [1] [4] [2] [8] [9]. The position of this survey is therefore that automated vacuolation quantification in fish is currently validated at the level of biological plausibility, one rung below the reader-agreement standard now demanded in clinical computational pathology, and two rungs below the external generalisation standard that reviewers of that field identify as still largely unmet [10].

The single aquatic pipeline validates by reproducing a prior ordinal call, not by measuring segmentation accuracy

The flounder study quantified the proportion of liver tissue affected by hepatocellular vacuolation across two estuaries and reported 6.88% (sd 3.01) in Seine-versus 5.46% (sd 2.86) in Canche fish, stating explicitly that this result "is coherent with the previous qualitative histological assessment done on the same histological samples" and that this coherence "helps validate the method" [6]. That is a directional concordance argument: the algorithm is validated by recovering a difference a pathologist had already seen. No Dice coefficient, per-object detection rate, held-out annotated region, inter-reader benchmark, or rescan repeatability statistic is reported in the abstract [6], in contrast to segmentation studies elsewhere that report model-expert overlap and specificity against expert annotation [3] or benchmark against independent pathologist annotations and compare performance to inter-rater agreement itself [13]. The same study also demonstrates why such a pipeline matters — the multipollutant mixture effect on vacuolation burden was detectable in one estuary and not the other, which requires a continuous outcome that ordinal grading cannot supply [6].

In MASH pipelines, validation strength tracks feature difficulty: vacuolation validates well, and the features that make disease "steatohepatitis-like" validate worst

The clearest quantitative evidence for this gradient is a mouse-plus-human nnU-Net pipeline whose agreement with conventional histology scores falls monotonically with semantic complexity: fat area R=0.91, inflammation density R=0.70, hepatocellular ballooning density R=0.49 [1]. Regional AI steatosis quantification correlated with manual grading at rs=0.72 while inflammation cell counts correlated at rs=0.71 in portal regions but effectively not at all in lobular regions (rs<0.1) [11]. Trial-scale machine learning features for steatosis, lobular inflammation, ballooning and fibrosis were "significantly correlated" with pathologist CRN grades and stages, and were then carried further into association with non-invasive biomarkers, a stellate-cell gene signature and liver-related event risk [7]. The pattern matters for aquatic work: a fish pipeline optimised on vacuoles is validating the easiest target in the panel, so its success says little about its capacity for the ballooning-like and inflammatory changes that would justify a "NASH-like" label.

Where rigorous designs exist, the reference standard is reader consensus — and that ceiling is low

Foundation-model classifiers for macrovesicular steatosis reached 96.4% accuracy at the clinically relevant <30% versus ≥30% threshold against a ground truth built from independent review of permanent sections by two liver pathologists, versus 84.0% for on-call surgical pathologists, with the difference non-significant; at five-category resolution performance collapsed to 57.1% for the best model and 58.7% for pathologists, with errors concentrated in the 5–<30% band [4]. Donor-liver benchmarking was motivated precisely by "substantial inter- and intra-observer variability" in quantifying total and macrovesicular steatosis and ballooning, and released expert annotations for those features as a public benchmark [2]. The most explicitly regulatory-shaped design located defined non-inferiority and superiority against pathologist consensus with prespecified agreement margins, and separately established greater than 99% repeatability for every scoring metric examined [8]. Independent work confirms how much headroom reader variability creates: manual crypt counts deviated from the group mean by 34±26% between individuals, falling to 11±6% only among the three most experienced counters, while the automated workflow deviated from the expert mean by 10±8% [20]. Two consequences follow: reader consensus is a soft reference standard, and study size can decide the verdict outright — a transfer-learned liver classifier failed at 42.86% test accuracy where the liver dataset fell below the >100–200 sample thresholds the authors derive, while the same approach reached 100% on a larger gastrointestinal set [5].

Orthogonal-modality and time-course agreement is the strongest validation argument available, and it has no aquatic analogue yet

Two designs move beyond correlating with a pathologist: the qHisto pipeline's fat area was correlated with quantitative MRI hepatic fat fraction (R=0.87), giving an independent physical measurement of the same biology [1], and a murine ductular reaction model was validated against a multiple linear regression of immunohistochemically quantified epithelial and stromal components (R²=0.88) and against published spatial transcriptomics of the same injury model [3]. A semi-automated alveolar morphometry method likewise selected parameters on a test image set against mean linear intercept and then validated on a separate dataset [19]. These are the templates a fish vacuolation study could adopt — hepatic triglyceride content, whole-body lipid, or lipid-droplet size distributions as the orthogonal anchor — but no retrieved aquatic study does so.

Transfer to non-mammalian tissue is unproven, and the retrieved veterinary evidence points toward failure rather than graceful degradation

A general-purpose multimodal model applied to poultry parasite microscopy reached 14.9% accuracy with a strong class bias and no correct calls for three of seven species, with segmentation only partial, leading the authors to conclude such models are "not yet reliable for standalone" diagnosis without domain-specific fine-tuning and expert validation [14]. Veterinary AI infrastructure is described as fragmented relative to human medicine, motivating a dedicated ecosystem of live benchmark datasets and evaluation frameworks [17]. Tooling is similarly mammal-centric: existing automated muscle morphometry tools are "largely tailored to mammalian models, with limited applicability to teleosts," which motivated a teleost-specific plugin [21], while zebrafish microCT morphometry required purpose-built automation to eliminate inter-observer variability and was benchmarked against manual and semi-automated alternatives [12].

Pre-analytic and cross-site robustness are the axes least tested anywhere, and most exposed in field-sampled fish

A meta-analysis of AI on whole slide images found mean sensitivity 96.3% and specificity 93.3% across 48 pooled studies, but all 100 included studies had at least one domain at high or unclear risk of bias, with study design substantially heterogeneous and dataset composition frequently undocumented [10]. Domain adaptation research explicitly targets inter-hospital variation in staining and digitisation protocols as a source of domain shift [16], and benchmarked artifact segmentation methods differed significantly in performance, with out-of-focus regions and dark spots hardest to segment [18]. Blur at levels typical of real-world slides did not significantly degrade prediction concordance in one large retrospective study, with small effect sizes (Cohen's d<0.2) [15]. Field-sampled wild fish introduce variability — multi-estuary sampling, differing fixation and sectioning histories — that no retrieved paper quantifies for a vacuolation model [6].

Conflicts and gaps

Three disagreements are unresolved. First, whether correlation with ordinal grade constitutes validation: three pipelines treat it as sufficient [1] [11] [6], whereas the reader-agreement and diagnostic-accuracy literature treats categorical scores as a reference standard requiring agreement, bias and repeatability statistics and flags correlation-only designs as high risk of bias [10] [8]. Second, whether AI matches or exceeds pathologists: one study found statistical parity in binary classification and no advantage in five-class grading [4], while another reported superiority on several subgrades and aggregate indices [8]. Third, and specific to this question, whether fish hepatocellular vacuolation is a lesion or a physiological storage state is never operationalised: the flounder study treats it as pathology whose extent tracks contamination [6], while mammalian work treats steatosis as a continuous burden with a clinically meaningful threshold [4] — but no retrieved paper establishes a normal reference range, validated cut-off, or biological anchor for the fish feature, so there is nothing for an algorithm's output to be accurate against. Also absent: any public annotated fish liver whole slide benchmark equivalent to the donor-liver release [2], and any cross-species fine-tuning experiment on liver tissue.

What would change the answer

A multi-reader annotated flounder or zebrafish liver whole slide set, released publicly, with segmentation scored by Dice against consensus annotations and reported alongside the human inter-rater Dice — so that the model is judged against the ceiling rather than against zero — would convert the field from plausibility to agreement validation, following the framing already used for vessel segmentation [13] and consensus-agreement design [8]. A repeatability and reproducibility study re-scanning, re-staining and re-sectioning the same blocks, reporting ICC and Bland–Altman limits of agreement, would establish the measurement error floor. A pre-registered external validation holding out sampling site, laboratory and scanner — the structure set out for prostate biopsies on an independent national register cohort [9] — would test the generalisation claim that field-sampled aquatic studies implicitly make. An orthogonal anchor — hepatic triglyceride or whole-body lipid content, or droplet-size distribution statistics as used to show that nuclear deformation is driven by large droplets rather than total fat [1] — would provide validation against biology rather than against a previous reader. Finally, a direct zero-shot versus fine-tuned comparison of pathology foundation models on fish liver, mirroring the poultry microscopy result [14] and the SSL-embedding benchmark [4], would establish whether human-trained encoders transfer to teleost tissue or whether, as the veterinary benchmark literature implies, dedicated aquatic datasets and training are required [17].

References

[1] Spirgath, K., Huang, B., Safraou, Y., Kraftberger, M., Dahami, M., Kiehl, R., Stockburger, C. H. F., Bayerl, C., Ludwig, J., Jaitner, N., Kühl, A., Asbach, P., Geisel, D., Hillebrandt, K. H., Wells, R. G., Sack, I., Tzschätzsch, H.. "Automated segmentation and quantification of histological liver features for MASH/MASLD scoring". medRxiv. 2026-02-15T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.13.26346163

[2] Liangrui Pan, Xingchen Li, Zhongyi Chen, Ling Chu, Shaoliang Peng. "DLiPath: A Benchmark for the Comprehensive Assessment of Donor Liver Based on Histopathological Image Dataset". arxiv. 2025-06-05T00:00:00.000Z. URL: https://arxiv.org/abs/2506.03185

[3] Feeley, N., Williams, K., Field, D., McCaffrey, C., Davies, K., Warden, H. B., Boulter, L., Thorn, S., Wigmore, S., Harrison, E., Forbes, S., Tomlinson, I., Kendall, T. J., Guest, R. V.. "DUCK-Net: Automated deep learning segmentation of Ductular Reaction in murine liver injury captures multicellular niche dynamics from H&E morphology". bioRxiv. 2026-02-10T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.09.704844

[4] Koga, S., Guda, A., Wang, Y., Sahni, A., Wu, J., Rosen, A., Nield, J., Nandish, N., Patel, K., Goldman, H., Rajapakse, C., Walle, S., Kristen, S., Tondon, R., Alipour, Z.. "Accuracy of Foundation AI Models for Hepatic Macrovesicular Steatosis Quantification in Frozen Sections". medRxiv. 2025-09-17T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.16.25335833

[5] Adeluwoye, A. O., Gbadegesin, M. O., James, F. M., Otegbade, P. S., Alabetutu, A.. "Algorithm-Based Model for Gastrointestinal and Liver Histopathological Analysis Using VGG16 and Specialized Stains: Statistical Validation of Thresholds in AI-Driven Digital Pathology". medRxiv. 2026-07-28T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.08.26350456

[6] Geslin, V., Diop, M., Couteau, J., Amara, R., Gulsrud, T. O., Laroche, J.. "New whole slide image analysis method for detection and quantification of hepatocellular vacuolization in European flounder (Platichthys flesus) in multipollutant exposure context". bioRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.08.757637

[7] Gerardin, Y., Stanford-Moore, A., Patel, N., Zhang, Y., Zhang, J., Mistry, P., Kartik, D., Liu, Y., Indorf, N., Fahy, D., Singh, G., Glickman, J., Resnick, M., Windholz, L., Alkhouri, N., Billin, A., Watkins, T., Brosnan-Cashman, J. A., Jayson, C., Lee, J., Glass, B., Beck, A. H., Iyer, J., Drage, M. G., Murray, L., Egger, R.. "Quantitative fibrosis features from digitized H&E-stained liver biopsies reveal granular insights into fibrosis and disease progression in metabolic dysfunction-associated steatohepatitis: a retrospective analysis of STELLAR-3 and STELLAR-4". medRxiv. 2026-06-02T00:00:00.000Z. URL: https://doi.org/10.1101/2025.06.12.25328580

[8] Tahir, W., Shamshoian, J., Tauber, J., Clinton, L. K., Griffin, M., Shah, C., Singh, G., Fahy, D., Sucipto, K., Brosnan-Cashman, J., Altepeter, T. A., Bhattacharya, S., Crandall, W., Duan, C., Gale, J. D., Gupta, V., Haarmann, H., Harpaz, N., Hooper, A. T., Horowitz, J., Hurtado-Lorenzo, A., Hussaini, B. E., Jairath, V., Jones, A., Kostiuk, B., Laroux, F. S., Lissoos, T., McBride, R. B., Najdawi, F., Nayyar, A., Osterman, M. T., Panchal, P., Ruane, D., Travis, S., Wilson, L., Jayson, C.. "Standardized, repeatable ulcerative colitis histology scoring and endpoint assessment using an automated, foundation model-based tool". medRxiv. 2026-07-09T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.09.26355212

[9] Ji, X., Zelic, R., Aspegren, O., Mulliqi, N., Fiorentino, M., Giunchi, F., Molinaro, L., Boman, S. E., Szolnoky, K., Liu, L. X., Pettersson, A., Vincent, P. H., Eklund, M., Akre, O., Kartasalo, K.. "Retrospective Validation of an Artificial Intelligence System for Diagnostic Assessment of Prostate Biopsies on the ProMort Cohort: Study Protocol". medRxiv. 2025-09-23T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.22.25336169

[10] Clare McGenity, Emily L Clarke, Charlotte Jennings, Gillian Matthews, Caroline Cartlidge, Henschel Freduah-Agyemang, Deborah D Stocken, Darren Treanor. "Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy". arxiv. 2025-05-23T00:00:00.000Z. URL: https://doi.org/10.1038/s41746-024-01106-8

[11] Windell, D., Magness, A., Beyer, C., Thomaides-Brears, H., Larkin, S., Hobson, K., Aljabar, P., Fleming, K., Fryer, E., Kendall, T., Kainth, R., Wakefield, P., Langford, C., Bedossa, P., Goldin, R.. "AI portal tract detection and characterisation for a regional analysis of steatosis and inflammation in MASLD, MASH, and AIH". medRxiv. 2025-04-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.23.25326290

[12] Kanani, M. M., Kwon, R. Y.. "Automatic Body Morphometric Analysis of Adult Zebrafish Using MicroCT". bioRxiv. 2025-09-21T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.19.677450

[13] Grion, G., Hussain, R., Colella, F. E., Roufail, K., Uccella, S., Frapolli, R., Matteo, C., Mintemur, O., Pennati, F., Renne, S. L.. "Vessel Spatial Analysis (VeSpA): a tool for whole slide image segmentation, morphometry, and QuPath extension.". bioRxiv. 2026-06-20T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.15.732366

[14] Ali Alsalama, Ahmed Kubba, Manar Abu Talib. "Toward AI-Assisted Poultry Coccidiosis Diagnosis: Evaluating Gemini and BiomedParse on Eimeria Microscopy Images". arxiv. 2026-09-29T00:00:00.000Z. URL: https://arxiv.org/abs/2609.31679

[15] Kim, H. H., Ko, Y. S., Kim, K.. "Impact of Whole Slide Image Blurriness on the Robustness of Artificial Intelligence in Real World Setting: Retrospective Observational Study". medRxiv. 2025-03-06T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.05.25323268

[16] Ilán Carretero, Pablo Meseguer, Rocío del Amor, Valery Naranjo. "Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation". arxiv. 2024-12-06T00:00:00.000Z. URL: https://arxiv.org/abs/2412.0426

[17] Katherine J. Koebel, Renata Ivanek, Parminder Basran, Jennifer Sun, Casey Cazer, Gayatri Anil, Tommaso Banzato, Liz Barton, Dottie Cimino Brown, Leszek Bukowski, Marta Castelhano, Simon Coghlan, Christopher De Sa, Nikki Dutta, Hala Ghali, Julio Giordano, Joshua Glass, David C. Hogg, Nicholas Hollingshead, Adam Kiehl, Ziang Liu, Monica Meduri, Param Mohapatra, Kim Nayyer, Amelia Nelson, Peter-John Mantyla Noble, Christopher Pinard, Beryl Rabindran, Scott Ross, Ron Seccia, Moira Sheehan, Fei Wang, Tracy L. Webb, Kilian Q. Weinberger, Heather M. Whitney, Chen Xin. "VETNET: Assembling Benchmark Datasets for AI-Driven Veterinary Innovation". arxiv. 2026-10-14T00:00:00.000Z. URL: https://arxiv.org/abs/2610.14149

[18] Julia Merta, Daniel Mendelsohn, Benedikt Schoser, Joanna Polanska, Michal Marczyk. "Benchmarking and Improvement of Artifact Segmentation Methods in H&E-Stained Whole Slide Imaging". arxiv. 2026-10-20T00:00:00.000Z. URL: https://arxiv.org/abs/2610.24037

[19] Gainer, D. J., Ormiston, M. L.. "An automated morphometric approach to evaluate distal lung patterning in mouse models of Bronchopulmonary Dysplasia". bioRxiv. 2025-09-17T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.11.675692

[20] Baikalov, A., Wang, E., Neill, D., Shetty, N., Waldrop, T., Liu, K., Delahousessaye, A., Aguilar, E., Mims, N., Bartzsch, S., Schueler, E.. "Modernizing histopathological analysis: a fully automated workflow for the digital image analysis of the intestinal microcolony survival assay". bioRxiv. 2024-12-12T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.09.627578

[21] Lu, Y., Pan, M., Jamwal, V., Locop, J., Ruparelia, A. A., Currie, P. D.. "fishROI: A specialized workflow for semi-automated muscle morphometry analysis in teleosts". bioRxiv. 2026-03-30T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.27.714781

How this survey was made

This survey answers one question: “How are deep learning and whole slide image analysis pipelines being validated for automated quantification of hepatic vacuolation and other non-alcoholic steatohepatitis-like histopathology in fish and other non-mammalian aquatic model species?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 21 papers whose work this text draws on, chosen from the 121 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 21, 2026 and covered work from the previous two years.