Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data Analytics & AI · ML computational pathology

Machine Learning in Computational Pathology for the Pharmaceutical Industry

Computational pathology is the most industrially useful application of machine learning in histology, because the pharmaceutical use cases (biomarker stratification, trial enrichment, consistent central reading and prognostic endpoints) require reproducibility more than novelty. The foundation models now being trained on over a million slides are capable, yet their reported operating points indicate that an area under the curve (AUC) of 0.95 and a usable screening tool are not equivalent.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time16 minutes Sources59 peer-reviewed

The case for machine learning in pathology does not rest on reading slides better than pathologists. It rests on reading them identically every time, at a scale that no reading panel can match, and on extracting quantitative features from morphology that cannot be graded by eye. For a sponsor running a trial across forty sites, this reproducibility is the principal benefit. The PANDA challenge provides the clearest demonstration of current capability: algorithms built by 1,290 developers on 10,616 digitised prostate biopsies reached a quadratically weighted kappa of 0.862 against expert uropathologists on US external validation and 0.868 on European external validation.

Learning objectives

  • Why whole-slide images present a distinctive machine learning problem, and how weakly supervised learning addresses it
  • What the challenge datasets established, and how the PANDA agreement figures should be interpreted
  • The scale and reported performance of the current pathology foundation models
  • Where a pharmaceutical sponsor obtains concrete value, stage by stage
  • Why a high AUC and a deployable operating point are different things
  • What a computational pathology study must report to be credible to a regulator or a reviewer

Common misconceptions

  • “The value of computational pathology is beating the pathologist.” The value lies in reproducibility and throughput. Inter-observer variation in grading is the problem that an industrial sponsor faces, and an algorithm that performs at expert level consistently solves it — which is what the PANDA challenge’s external-validation agreement of 0.862 and 0.868 against expert uropathologists demonstrates.
  • “A pan-cancer AUC of 0.95 means the model is ready to screen specimens.” The operating point is decisive. Vorontsov and colleagues report Virchow at a specimen-level AUC of 0.950 pan-cancer and 0.937 on rare cancers, with a specificity of 72.5% at 95% sensitivity. In a specimen screening workload, approximately a quarter of negative specimens would therefore still be flagged.
  • “Whole-slide image analysis is image classification with bigger files.” A whole-slide image is gigapixel-scale and its label usually applies to the whole slide rather than to any region, so the supervision signal is weak by construction. Campanella and colleagues established clinical-grade performance in Nature Medicine using weakly supervised learning on whole slides for this reason, and Lu and colleagues’ data-efficient approach in Nature Biomedical Engineering is built around the same constraint.
  • “Foundation models remove the need for task-specific evaluation.” They reduce the labelled data that each task requires but do not transfer their validation. PathOrchestra was trained on 287,424 whole-slide images and then evaluated on 112 downstream tasks using 27,755 slides and about 9.4 million region-of-interest images; evaluation remains the substantive work and must be conducted task by task.
  • “Digital pathology performance is a property of the algorithm.” It is a property of the algorithm, the stain, the scanner and the laboratory. Stain protocol, section thickness, fixation and scanner optics all vary between sites, and a model trained at one laboratory has partly learned the characteristics of that laboratory, which is why the Digital Pathology Association white paper by Abels and colleagues in The Journal of Pathology treats preanalytical standardisation as part of the definition of the field.

Definition of computational pathology

Computational pathology is the quantitative analysis of digitised tissue specimens, usually whole-slide images, to produce diagnostic, prognostic or predictive outputs. Abels and colleagues’ Digital Pathology Association white paper in The Journal of Pathology is the reference statement of definitions, best practices and recommendations for regulatory guidance, and it merits reading before any project begins because it places the preanalytical chain within the scope of the discipline rather than before it 46. Madabhushi and Lee’s review in Medical Image Analysis remains the clearest statement of the challenges and opportunities.

Three distinctions are important when planning work in this area. Detection and grading tasks replicate an assessment that a pathologist already performs, so a reference standard exists and the question is one of agreement. Molecular inference tasks predict a property that the slide was not previously thought to contain, such as mutation status or expression subtype, so the reference standard derives from a different assay and the question is whether morphology carries the information. Prognostic tasks predict outcome directly from morphology, where the reference standard is follow-up and the question is whether the model has found features no grading system encodes.

These are not degrees of difficulty on a single scale; they are different claims requiring different validation. A grading model is validated against adjudicated expert reads. A molecular inference model is validated against the orthogonal assay, which has its own error rate. A prognostic model is validated against outcomes in an independent cohort with adequate follow-up, and must be shown to add to established prognostic factors rather than recapitulate them. Conflating the three is the most common structural weakness in manuscripts in this area. See deep learning challenges in medical imaging for the shared methodology.

Whole-slide images and weakly supervised learning

A whole-slide image is very large, and its label is usually attached to the entire slide. “This biopsy contains carcinoma” gives no indication of location, so a conventional supervised pipeline has no target from which to learn at the resolution at which the signal resides.

The standard response in the field is weakly supervised, tile-based learning: the slide is divided into tiles, tile-level representations are learned, and these are aggregated into a slide-level prediction under the slide-level label. Campanella and colleagues demonstrated clinical-grade computational pathology in this way in Nature Medicine; Lu and colleagues’ data-efficient and weakly supervised approach in Nature Biomedical Engineering made the approach practical with far fewer slides; Laleh and colleagues benchmarked the available pipelines systematically in Medical Image Analysis, which should be consulted before a pipeline is chosen 24,36,45. Wang and colleagues apply transformer-based unsupervised contrastive learning to histopathological classification in the same journal, and Krishnan, Rajpurkar and Topol review self-supervised learning in medicine more broadly in Nature Biomedical Engineering 31,33.

The alternative is dense annotation, which is expensive but occasionally justified. HoVer-Net, developed by Graham and colleagues and reported in Medical Image Analysis, performs simultaneous segmentation and classification of nuclei across tissue types; Greenwald and colleagues achieved human-level whole-cell segmentation in Nature Biotechnology through large-scale annotation 34,48. Where a downstream analysis requires cell-level quantification (counts, spatial relationships or density gradients), tile-level aggregation will not supply it and the cost of annotation must be accepted.

Two consequences follow for study design. First, the aggregation step is a modelling choice with substantive consequences and should be reported rather than obscured: how tile predictions become a slide prediction determines what the model is sensitive to. Second, class imbalance in histopathology is severe, since most tiles on most slides are uninformative; Escobar Diaz Guerrero and colleagues set out an augmentation methodology to address it in the Journal of Imaging Informatics in Medicine, and whatever correction is applied alters the meaning of the output probabilities and must therefore be stated.

Findings established by challenge datasets

Computational pathology is unusual in having a set of public challenges that produced comparable results, and these provide an appropriate basis for calibrating expectations.

The CAMELYON16 assessment by Ehteshami Bejnordi and colleagues in JAMA evaluated deep learning algorithms for detecting lymph node metastases in women with breast cancer; Bándi and colleagues extended it in CAMELYON17 from detecting individual metastases to classifying lymph node status at the patient level, in IEEE Transactions on Medical Imaging 52,55. TUPAC16, by Veta and colleagues, addressed the prediction of breast tumour proliferation from whole-slide images, and the GlaS contest, by Sirinukunwattana and colleagues, addressed gland segmentation in colon histology, both in Medical Image Analysis.

PANDA is the most informative challenge for industrial purposes because of its scale and its external validation design. Bulten and colleagues, in Nature Medicine, report a challenge involving 1,290 developers and 10,616 digitised prostate biopsies, in which algorithms reached a quadratically weighted kappa of 0.862 (95% CI 0.840–0.884) against expert uropathologists on US external validation and 0.868 (95% CI 0.835–0.900) on European external validation 23.

These figures merit precise interpretation, because they are often read as accuracy figures. Quadratically weighted kappa measures agreement on an ordinal scale, with larger disagreements penalised more heavily, which makes it the appropriate metric for Gleason grading, where a discrepancy of one grade matters much less than one of three. Agreement in the high 0.8s against expert uropathologists, maintained across two independent external validation sets on two continents, is a strong result; it is, however, a statement about agreement with a reference reading, not about clinical outcome 23. The work of Huang and colleagues in Nature Communications, in which a deep neural network trained on gigapixel images improved lymph node metastasis detection in clinical settings, is the closest approximation to the next step 27.

The public challenges and what each established

Challenge and sourceTaskWhat it established
CAMELYON16 — Ehteshami Bejnordi and colleagues, JAMADetection of lymph node metastases in breast cancerThat algorithms could be assessed diagnostically against pathologists on a common dataset
CAMELYON17 — Bándi and colleagues, IEEE Transactions on Medical ImagingPatient-level lymph node status classificationThe move from detecting individual metastases to a patient-level clinical determination
TUPAC16 — Veta and colleagues, Medical Image AnalysisPredicting breast tumour proliferation from whole-slide imagesThat a quantitative biological property could be predicted from morphology at slide level
GlaS — Sirinukunwattana and colleagues, Medical Image AnalysisGland segmentation in colon histologyA common standard for a structural segmentation task underpinning downstream analysis
PANDA — Bulten and colleagues, Nature MedicineGleason grading of prostate biopsiesExpert-level agreement sustained on two independent external validation sets: quadratically weighted kappa 0.862 (95% CI 0.840–0.884) in the US and 0.868 (95% CI 0.835–0.900) in Europe, from a challenge involving 1,290 developers and 10,616 digitised prostate biopsies

As reported in each publication. Quadratically weighted kappa measures agreement on an ordinal scale with larger disagreements penalised more heavily — the appropriate metric for grading, and a measure of agreement with a reference reading rather than of clinical benefit.

0.862quadratically weighted kappa against expert uropathologists, US external validation (95% CI 0.840–0.884)
0.868the same, European external validation (95% CI 0.835–0.900)
1,290developers who entered
10,616digitised prostate biopsies

Expert-level grading agreement, holding across two independent external validation sets. Figures from Bulten W, Kartasalo K, Chen PHC, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nature Medicine.

Foundation models and the benefits of scale

The current generation of pathology models is pretrained on slide collections large enough to have altered the economics of the field: a sponsor no longer needs to assemble a task-specific training cohort of thousands of annotated slides to obtain a usable starting point.

The reported scale is substantial. Xu and colleagues’ Prov-GigaPath, in Nature, was pretrained on 171,189 whole slides from more than 30,000 patients across 31 tissue types, amounting to about 1.3 billion image tiles (1,384,860,229) 16. Vorontsov and colleagues’ Virchow, in Nature Medicine, is a 632-million-parameter model trained on 1,488,550 slides from 119,629 patients, reporting a specimen-level AUC of 0.950 pan-cancer and 0.937 on rare cancers 15. Ding and colleagues’ TITAN, in the same journal, was pretrained on 335,645 whole-slide images and reported a 56.52% balanced-accuracy gain over PRISM on zero-shot multiclass classification 4. Yan and colleagues’ PathOrchestra, in npj Digital Medicine, was trained on 287,424 whole-slide images and evaluated on 112 downstream tasks using 27,755 slides and about 9.4 million region-of-interest images 13. Yang and colleagues’ BEPH, in Nature Communications, was pretrained on about 11 million unlabelled TCGA patches and reported 94.05% patient-level accuracy on BreakHis and a macro-AUC of 0.994 for renal cell carcinoma subtyping 3.

Three observations apply to this set of results. First, the headline figures are benchmark figures, and the benchmarks differ between models, so the numbers above do not constitute a ranking despite their appearance. Second, the rare-cancer result for Virchow (0.937, against 0.950 pan-cancer) is the most practically relevant figure in the group, because rare disease is where the individual experience of a pathologist is most limited and where a model trained across 119,629 patients has the clearest structural advantage. Third, the zero-shot gains reported for models such as TITAN are measured against another model rather than against a clinical standard; a 56.52% improvement in balanced accuracy over a predecessor indicates that the field is advancing rapidly but says nothing about whether either model is adequate for a given task 4.

Pathology foundation models: reported pretraining scale and headline results

Model and sourcePretraining scaleReported performanceNote
Prov-GigaPath — Xu and colleagues, Nature171,189 whole slides from more than 30,000 patients across 31 tissue types; about 1.3 billion image tiles (1,384,860,229)—Whole-slide foundation model built from real-world data
Virchow — Vorontsov and colleagues, Nature Medicine1,488,550 slides from 119,629 patients; 632 million parametersSpecimen-level AUC 0.950 pan-cancer and 0.937 on rare cancers; 72.5% specificity at 95% sensitivityThe rare-cancer figure is the most practically relevant; the operating point is the figure relevant to deployment
TITAN — Ding and colleagues, Nature Medicine335,645 whole-slide imagesBalanced-accuracy gain of 56.52% over PRISM on zero-shot multiclass classificationGain is measured against another model, not a clinical standard
PathOrchestra — Yan and colleagues, npj Digital Medicine287,424 whole-slide imagesEvaluated on 112 downstream tasks using 27,755 slides and about 9.4 million region-of-interest imagesThe breadth of evaluation is the contribution, and it shows that each task requires its own assessment
BEPH — Yang and colleagues, Nature CommunicationsAbout 11 million unlabelled TCGA patches94.05% patient-level accuracy on BreakHis; macro-AUC 0.994 for renal cell carcinoma subtypingBenchmark-specific results on public datasets

Figures exactly as reported in each publication. These are not comparable as a ranking: the pretraining corpora, evaluation benchmarks, tasks and metrics differ between models, and several of the headline figures are relative to a predecessor model rather than to a clinical standard.

Value for pharmaceutical sponsors

The applications that deliver value are those in which reproducibility or scale is the binding constraint, not those in which the algorithm is most impressive.

Biomarker status from morphology. Coudray and colleagues, in Nature Medicine, showed classification and mutation prediction from non-small cell lung cancer histopathology images; Kather and colleagues, in Nature Cancer, demonstrated pan-cancer image-based detection of clinically actionable genetic alterations; Gamble and colleagues, in Communications Medicine, determined breast cancer biomarker status and the associated morphological features 42,51. For a sponsor, the immediate application is pre-screening, that is, using morphology to prioritise which specimens proceed to molecular testing, which reduces assay expenditure without replacing the assay.

Trial enrichment and stratification. A model that identifies a morphologically defined subpopulation can enrich a trial, provided that the model is locked and validated before the trial rather than developed on it. The prognostic literature is relevant here: Skrede and colleagues, in The Lancet, reported deep learning for prediction of colorectal cancer outcome in a discovery and validation study; Wulczyn and colleagues reported interpretable survival prediction for colorectal cancer in npj Digital Medicine and deep learning survival prediction across multiple cancer types in PLOS ONE; Lee and colleagues derived prognostic contextual histopathological features via graph deep learning in Nature Biomedical Engineering 28,38,39,40.

Central reading at scale, with consistency. This is the least conspicuous and most reliable source of value. The PANDA external-validation agreement figures are the evidence that grading can be performed reproducibly at expert-level agreement, and reproducibility across sites is precisely what a multi-centre trial cannot otherwise guarantee 23.

Tissue-based mechanism and immune phenotyping. Saltz and colleagues, in Cell Reports, mapped spatial organisation and molecular correlation of tumour-infiltrating lymphocytes using deep learning; Lu and colleagues characterised TILs in breast cancer from histopathology and multiomics data in JCO Clinical Cancer Informatics; Ghahremani and colleagues inferred multiplex immunofluorescence for immunohistochemical quantification in Nature Machine Intelligence; Chen and colleagues performed pan-cancer integrative histology-genomic analysis via multimodal deep learning in Cancer Cell; Beck and colleagues’ much earlier work in Science Translational Medicine found stromal features associated with survival, which is the origin of this entire line of enquiry 26,30,41,53,59.

Safety and specialised reads. Lipkova and colleagues, in Nature Medicine, demonstrated deep learning-enabled assessment of cardiac allograft rejection from endomyocardial biopsies, an example of a low-volume, high-variability read in which consistency has evident value 25. Lu and colleagues’ AI-based prediction of origins for cancers of unknown primary in Nature is the equivalent for a diagnostic problem that routine work-up cannot resolve 35. Liu and colleagues’ work on non-destructive 3D pathology in Nature Biomedical Engineering indicates how the specimen itself may change. See systematic review of machine learning in the pharmaceutical industry and AI and ML use cases in clinical development.

Where machine learning in pathology pays, by stage

StageUse caseWhy machine learning helps hereEvidence base
Discovery and translationalTissue-based mechanism, immune phenotyping, spatial biologyQuantifies morphology and spatial relationships no grading system encodesSaltz and colleagues (Cell Reports); Chen and colleagues (Cancer Cell); Ghahremani and colleagues (Nature Machine Intelligence); Beck and colleagues (Science Translational Medicine)
Preclinical safetyToxicologic pathology reads at volumeHigh-volume, consistency-critical scoring; validation path set by the preclinical frameworkAdjacent AI toxicology literature: Lin and Chou and Chou and Lin (Toxicological Sciences); Zhang and colleagues (Briefings in Bioinformatics)
Biomarker and companion diagnostic developmentPredicting molecular status from morphology to pre-screen for assayReduces assay spend by prioritising specimens, without replacing the assayCoudray and colleagues (Nature Medicine); Kather and colleagues (Nature Cancer); Gamble and colleagues (Communications Medicine)
Trial design and enrichmentMorphologically defined subpopulation selection; prognostic stratificationIdentifies strata reproducibly, provided the model is locked before the trialSkrede and colleagues (The Lancet); Wulczyn and colleagues (npj Digital Medicine and PLOS ONE); Lee and colleagues (Nature Biomedical Engineering)
Trial conductCentral reading and grading across sitesExpert-level agreement delivered identically at every site, a constraint that a multi-centre trial cannot otherwise meetPANDA challenge: Bulten and colleagues (Nature Medicine); CAMELYON16 and 17
Specialised and low-volume readsAllograft rejection, cancer of unknown primaryConsistency where individual reader experience is most limitedLipkova and colleagues and Lu and colleagues (Nature Medicine and Nature)

The use cases represent Pubrica’s framing; the evidence column lists the published demonstrations cited in this article. Inclusion indicates that a capability has been demonstrated in research, not that a validated product exists for that use.

Applications in preclinical research and drug discovery

Computational pathology sits alongside a much larger machine learning effort in discovery and preclinical development, and the connection merits attention because sponsors increasingly fund the two together.

Pun and colleagues, in Nature Reviews Drug Discovery, review target identification and assessment in the era of AI. Catacutan and colleagues cover machine learning in preclinical drug discovery in Nature Chemical Biology, and Sadybekov and Katritch survey computational approaches streamlining drug discovery in Nature 17,19. Pandey and colleagues, in Nature Machine Intelligence, describe the transformational role of GPU computing and deep learning in the same area, the computational advances that also made whole-slide foundation models possible.

On the safety side, Zhang and colleagues review applications of artificial intelligence in ADMET and toxicity prediction in Briefings in Bioinformatics; Lin and Chou cover machine learning and artificial intelligence in toxicological sciences, and Chou and Lin cover their use in physiologically based pharmacokinetic modelling, both in Toxicological Sciences 7,21,29. Ma and colleagues, in npj Digital Medicine, examine AI-driven virtual cell models in preclinical research, including their validation mechanisms and potential for clinical translation, which is the most speculative of these directions and the most candid about its limitations 5.

The point of connection for a histology programme is the toxicologic pathology read: digital assessment of treatment-related findings in preclinical studies is a high-volume, consistency-critical task of precisely the kind to which computational pathology is suited. It is also a task for which regulatory expectations are set by the preclinical framework rather than the diagnostic one, so the validation path differs from that of a clinical biomarker. Nonclinical study reporting and pharmaceutical industry services are where that work lands.

Limitations of headline performance figures

Four limitations apply, each of which should appear in the limitations section of a manuscript and often does not.

The operating point is the decisive figure. The specimen-level AUC of 0.950 reported for Virchow is excellent, and the reported specificity of 72.5% at 95% sensitivity is the figure on which a laboratory director would base any decision 15. These two numbers derive from the same model in the same paper, yet they support quite different impressions. Sensitivity and specificity at a prospectively chosen operating point should always be reported alongside any AUC.

Agreement with a reference read is not clinical benefit. PANDA’s kappa figures of 0.862 and 0.868 measure agreement with expert uropathologists 23. This is the appropriate measure for a grading task, but it is not evidence that patients managed using the algorithm have better outcomes than patients managed by a pathologist.

Preanalytical variation is part of the model. Stain protocol, fixation, section thickness, scanner optics and compression all vary between laboratories, and a model trained at one site has partly learned the characteristics of that site. This is the histological counterpart of the scanner-generalisation problem, and it is why validation on slides prepared and scanned at an independent laboratory is the minimum credible external test, rather than a different cohort drawn from the same equipment 46,57.

Benchmark performance is not task performance. The foundation model figures above are reported on different benchmarks, and the structure of PathOrchestra demonstrates the point in reverse: 112 downstream tasks were evaluated on 27,755 slides, because a general-purpose model must be re-evaluated for each use 13. Any claim that a foundation model is suitable for a particular task requires evaluation by the user on the user’s own data. The account by Kelly, Karthikesalingam, Suleyman, Corrado and King in BMC Medicine of what stands between algorithmic performance and clinical impact applies here with little modification, as does the Pubrica page on the challenges of delivering clinical impact with machine learning 49.

0.950specimen-level AUC, pan-cancer
0.937specimen-level AUC, rare cancers
72.5%specificity at 95% sensitivity

The same model, the same paper: the discrimination figure and the deployable operating point give different impressions. Figures from Vorontsov E, Bozkurt A, Casson A, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine.

Validation and regulatory pathway

A computational pathology tool intended for use rather than publication must satisfy a different reader, and the requirements are more specific than general AI guidance suggests.

The Digital Pathology Association white paper by Abels and colleagues is the starting document, since it was written partly to inform regulatory guidance and sets out best practices with that audience in view 46. The requirements that follow from it and from the general reporting standards converge on a short list: the intended use stated at the level of specimen type, stain, scanner and clinical question; a locked model version, with any subsequent change treated as a change to the device; validation on specimens prepared and digitised independently of the development pipeline; performance reported at a fixed operating point with calibration; and performance reported by site, by scanner and by relevant subgroup rather than pooled.

Reporting standards should be matched to the study type, as in any clinical AI work: TRIPOD+AI for the development and validation of a prediction model; STARD-AI for a diagnostic accuracy evaluation; CONSORT-AI, with SPIRIT-AI for the protocol, for a randomised trial in which the tool is the intervention; DECIDE-AI for an early-stage clinical evaluation within a reading workflow, which for pathology is where the practical questions arise, since a tool that changes what a pathologist does has effects that offline accuracy cannot show; and PROBAST+AI for the appraisal of other investigators’ models in a review. Reporting checklist review and regulatory affairs and causal inference in machine learning cover this ground.

Reporting a computational pathology study

The characteristic failure in these manuscripts is specific and recurrent: a strong result is described in a way that makes it impossible to determine whether it would hold in another laboratory. Four measures address most of this problem.

The specimen chain should be specified, not only the dataset. This includes the source of the tissue, how it was fixed and stained, the protocol, the scanner and magnification, and the compression used. A reader in another laboratory must decide whether their slides resemble those in the study, and no other information in the paper permits this judgement.

Data should be separated by laboratory, not by patient alone. Patient-level separation between training and test sets is necessary but insufficient. Validation on material prepared and digitised at a site that contributed nothing to development is what distinguishes a model that has learned the disease from one that has learned the stain.

The aggregation and the operating point should be reported. This includes how tile-level outputs were converted into a slide-level or patient-level prediction, and the threshold used and when it was chosen. These are the two decisions most likely to be unreported and most likely to matter.

The type of claim being made should be specified. The claim may be agreement with a reference read, inference of a molecular property or prediction of outcome, and it should be validated against the standard appropriate to it. For a prognostic claim, the model should be shown to add to established factors rather than restate them, because a morphology model that recovers stage and grade is of interest but not useful. Biostatistics support and bias and fairness review are where we check these.

The validation paragraph of a computational pathology methods section

Would not survive another laboratory

Whole-slide images from 1,100 cases were obtained from our institutional digital pathology archive. Cases were split 80:20 into training and test sets at the patient level. A weakly supervised deep learning model was trained on tiled image patches and achieved an AUC of 0.94 on the held-out test set, demonstrating that the model can accurately identify the target feature in routine histology.

No preanalytical chain. Fixation, stain platform, scanner, magnification and compression are all absent, so no other laboratory can judge whether their slides resemble these.

Single-source data with a patient-level split. Patient-level separation prevents the most basic form of leakage but does not address a model that has learned the stain or the scanner.

No external laboratory. The held-out set was prepared and digitised on the same equipment as the training set, so the transfer question is untested rather than answered.

Aggregation unreported. How tile predictions became a slide prediction is a modelling decision that determines what the model responds to, and it is not reported.

AUC alone, no operating point, no calibration. Nothing is stated about how the model would behave in use, and the specificity at a usable sensitivity is not shown.

“Can accurately identify the target feature in routine histology” generalises to all routine histology from a single archive and a single scanner.

Validated against the preanalytical chain

Whole-slide images from 1,100 cases were obtained from two sources: 820 cases from our institutional archive, formalin-fixed and paraffin-embedded, stained with haematoxylin and eosin on a single automated platform and digitised on one scanner at 40× with the compression setting given in the Supplement; and 280 cases from an external laboratory using a different staining platform and scanner, which contributed nothing to model development. Development used only the institutional cases, split at patient level into training and tuning partitions with no patient in both; the external laboratory cases formed the held-out test set. Slides were tiled at the stated magnification and tile-level predictions were aggregated to a slide-level score by attention-weighted pooling, with the aggregation function fixed before test evaluation; the operating point was selected on the tuning partition and not revisited. On the external test set the model achieved an AUC of 0.89, with sensitivity and specificity at the fixed operating point and calibration reported in Table 3; the corresponding internal figures were AUC 0.94, and both are reported so the transfer gap is visible. Results are reported separately for each staining platform. The reference standard was an adjudicated read by two pathologists blinded to the model output with a third resolving disagreement, and we report agreement with that standard rather than any claim about clinical outcome.

The preanalytical chain is specified for both sources, which is the information another laboratory needs before anything else in the paper is useful to them.

The held-out set comes from an independent laboratory with different staining and scanning, which is the minimum credible external test in histology and the one that distinguishes disease from stain.

Internal and external performance are both reported, so the transfer gap is a stated result rather than something that a reader must suspect.

The aggregation function and the operating point are both named and both fixed before test evaluation, closing the two most common routes to an optimistically tuned result.

Sensitivity, specificity and calibration appear at that operating point, and results are broken out by staining platform rather than pooled.

The claim is limited to agreement with an adjudicated reference read, with an explicit statement that it is not a claim about outcome, which is the distinction that the three-claim taxonomy exists to enforce.

The second version reports a lower headline figure and is far more likely to be believed, adopted and reproduced. In this field, the external-laboratory result is the only figure on which a purchaser or a regulator will act, so reporting it is not a concession but the principal finding.

Key points

Key points

Machine learning justifies its place in pathology through reproducibility at scale rather than through outperforming pathologists, which is why it suits pharmaceutical use: biomarker pre-screening, trial enrichment, consistent central reading and prognostic endpoints all require the same read every time. The PANDA challenge (1,290 developers and 10,616 digitised prostate biopsies) showed a quadratically weighted kappa of 0.862 (95% CI 0.840–0.884) against expert uropathologists on US external validation and 0.868 (95% CI 0.835–0.900) in Europe, indicating expert-level agreement across two continents. Foundation models have changed the economics: Prov-GigaPath was pretrained on 171,189 slides from more than 30,000 patients across 31 tissue types; Virchow on 1,488,550 slides from 119,629 patients, reporting a pan-cancer AUC of 0.950 and 0.937 on rare cancers; and PathOrchestra on 287,424 slides, subsequently evaluated across 112 downstream tasks. However, Virchow’s 72.5% specificity at 95% sensitivity is the figure that determines deployment, agreement with a reference read is not clinical benefit, and the preanalytical chain (stain, scanner and laboratory) is part of the model. Validation requires slides prepared and digitised at a site that contributed nothing to development.

Application to your study

Application to your study

  1. The type of claim being made should be specified. The three claims are agreement with a reference read, inference of a molecular property and prediction of outcome. Each requires a different reference standard, and conflating them is the most common structural fault in these manuscripts.
  2. The complete specimen chain should be reported. Fixation, stain platform and protocol, section thickness, scanner and model, magnification and compression should be reported for development and for every validation set; without this information, the result cannot be used by others.
  3. Validation should be conducted at an independent laboratory, not only on independent patients. Validation slides should be prepared and digitised on equipment that contributed nothing to development, and internal and external performance should be reported side by side so that the transfer gap is visible.
  4. The aggregation function and the operating point should be stated and fixed before test evaluation. The report should state how tile predictions are aggregated into a slide or patient prediction, the threshold used, and when it was chosen.
  5. Sensitivity and specificity at that operating point, with calibration, should be reported alongside any AUC. A specimen-level AUC of 0.95 and a specificity of 72.5% at 95% sensitivity describe the same model and support different decisions.
  6. For a prognostic claim, incremental value over established factors should be demonstrated. A morphology model that recovers stage and grade is of interest, but one that adds information beyond them is useful, and only the latter supports a stratification strategy.

Frequently asked questions

The quantitative analysis of digitised tissue specimens — usually whole-slide images — to produce diagnostic, prognostic or predictive outputs. Abels and colleagues’ Digital Pathology Association white paper in The Journal of Pathology is the reference statement of definitions, best practices and recommendations for regulatory guidance, and it deliberately places the preanalytical chain within the scope of the field: stain protocol, fixation and scanning are part of the system, not preliminaries to it. Three distinct kinds of task fall under the term, and they require different validation: replicating a read that a pathologist already performs, where the standard is expert agreement; inferring a molecular property from morphology, where the standard is an orthogonal assay; and predicting outcome directly, where the standard is follow-up in an independent cohort.

On the best available evidence, at expert-level agreement sustained across continents. Bulten and colleagues report in Nature Medicine that in the PANDA challenge, involving 1,290 developers and 10,616 digitised prostate biopsies, algorithms reached a quadratically weighted kappa of 0.862 (95% CI 0.840–0.884) against expert uropathologists on US external validation and 0.868 (95% CI 0.835–0.900) on European external validation. Two points apply to these figures. Quadratically weighted kappa is an agreement measure on an ordinal scale that penalises larger disagreements more heavily, which makes it the appropriate metric for grading. Moreover, agreement with a reference read is not evidence of clinical benefit; it does not establish that patients managed using the algorithm have better outcomes.

Whole-slide images are gigapixel-scale, and their labels usually apply to the entire slide rather than to any particular region, so there is no target at the resolution at which the signal resides. The standard solution in the field is weakly supervised, tile-based learning: the slide is tiled, tile-level representations are learned, and these are aggregated into a slide-level prediction under the slide-level label. Campanella and colleagues established clinical-grade performance in this way in Nature Medicine and Lu and colleagues made it data-efficient in Nature Biomedical Engineering; Laleh and colleagues benchmarked the available pipelines in Medical Image Analysis. Two consequences should be reported: the aggregation step is a modelling choice that determines what the model is sensitive to, and class imbalance is severe because most tiles on most slides are uninformative.

A foundation model is a pretrained representation that reduces the amount of labelled data required for each downstream task, which alters the economics of building a tool. The reported scale is large: Prov-GigaPath was pretrained on 171,189 whole slides from more than 30,000 patients across 31 tissue types, about 1.3 billion image tiles; Virchow is a 632-million-parameter model trained on 1,488,550 slides from 119,629 patients; TITAN was pretrained on 335,645 slides; PathOrchestra on 287,424; and BEPH on about 11 million unlabelled TCGA patches. What these models do not offer is transferred validation. The design of PathOrchestra illustrates this, with 112 downstream tasks evaluated on 27,755 slides, because a general model must be assessed task by task. Any claim of suitability for a particular task requires evaluation on the user’s own data.

Value arises in five areas, and the common feature is that reproducibility or volume is the binding constraint. The first is pre-screening specimens for molecular assay, using morphology to prioritise which samples are tested; Coudray and colleagues, Kather and colleagues and Gamble and colleagues all demonstrated prediction of molecular status from histology. The second is trial enrichment and prognostic stratification, with the model locked before the trial. The third is central reading and grading across trial sites at consistent expert-level agreement, as evidenced by the PANDA external validation figures. The fourth is tissue-based mechanism and immune phenotyping, including spatial analysis of tumour-infiltrating lymphocytes. The fifth is specialised low-volume reads, such as allograft rejection assessment or cancer of unknown primary, where individual reader experience is most limited.

Histology slides are manufactured objects, and the manufacturing process varies. Fixation time, stain platform and protocol, section thickness, scanner optics, magnification and image compression all differ between laboratories, and a model trained on one laboratory’s material has partly learned that laboratory’s appearance. For this reason, patient-level separation between training and test sets, while necessary, does not constitute external validation in pathology: the held-out slides were prepared and digitised on the same equipment. The minimum credible external test is validation on material prepared and scanned at a site that contributed nothing to development, with results reported separately by staining platform and scanner rather than pooled, and with internal and external performance reported side by side so that the transfer gap is visible.

Not on its own, and the same papers usually contain the figure that answers the question properly. Vorontsov and colleagues report Virchow at a specimen-level AUC of 0.950 pan-cancer and 0.937 on rare cancers, and, at 95% sensitivity, a specificity of 72.5%. In a specimen screening workload, this means that approximately a quarter of negative specimens would still be flagged, an operational fact around which a laboratory director can plan, whereas an AUC is not. The general rule for writing and for reading is that an AUC summarises performance across all thresholds, whereas a deployed tool uses only one. Sensitivity and specificity should be reported at a prospectively chosen operating point, with calibration, and the prevalence in the intended-use setting should be stated.

The starting point is the Digital Pathology Association white paper by Abels and colleagues, which sets out best practices with regulatory guidance in view; the AI reporting standard should then be matched to the study type: TRIPOD+AI for the development and validation of a prediction model; STARD-AI for a diagnostic accuracy evaluation; CONSORT-AI, with SPIRIT-AI for the protocol, for a randomised trial in which the tool is the intervention; DECIDE-AI for an early-stage clinical evaluation within a reading workflow, which for pathology is where the practical questions arise, since a tool that changes what a pathologist does has effects that offline accuracy cannot show; and PROBAST+AI for the appraisal of other investigators’ models. Alongside any of these, the intended use should be reported at the level of specimen type, stain, scanner and clinical question, and the model version should be treated as locked.

References

  1. Farahani S, Hejazi M, Tabassum M, Di Ieva A, Mahdavifar N, et al. , Liu S, et al. Diagnostic performance of deep learning for predicting glioma isocitrate dehydrogenase and 1p/19q co-deletion in MRI: a systematic review and meta-analysis. European Radiology. 2026;36(2):1562-1591. doi:10.1007/s00330-025-11898-2 read 9 October 2026. doi.org/10.1007/s00330-025-11898-2
  2. Pun FW, Podolskiy D, Izumchenko E, Mortlock A, Oprea TI, Scheibye-Knudsen M, Fortney K, Morgen E, Ren F, et al. , Zhavoronkov A, et al. Target identification and assessment in the era of AI. Nature Reviews Drug Discovery. 2026;25(7):534-552. doi:10.1038/s41573-026-01412-8 read 9 October 2026. doi.org/10.1038/s41573-026-01412-8
  3. Yang Z, Wei T, Liang Y, Yuan X, Gao R, Xia Y, Zhou J, Zhang Y, et al. , Yu Z, et al. A foundation model for generalizable cancer diagnosis and survival prediction from histopathological images. Nature Communications. 2025;16(1):2366. doi:10.1038/s41467-025-57587-y read 9 October 2026. doi.org/10.1038/s41467-025-57587-y
  4. Ding T, Wagner SJ, Song AH, Chen RJ, Lu MY, Zhang A, Vaidya AJ, Jaume G, Shaban M, Kim A, et al. , Mahmood F, et al. A multimodal whole-slide foundation model for pathology. Nature Medicine. 2025;31(11):3749-3761. doi:10.1038/s41591-025-03982-3 read 9 October 2026. doi.org/10.1038/s41591-025-03982-3
  5. Ma C, Zhang H, Rao Y, Jiang X, Liu B, Sun Z, Song Z, Gao Y, Cui Y, Liu X, et al. , Li Z, et al. AI-driven virtual cell models in preclinical research: technical pathways, validation mechanisms, and clinical translation potential. npj Digital Medicine. 2025;9(1):25. doi:10.1038/s41746-025-02198-6 read 9 October 2026. doi.org/10.1038/s41746-025-02198-6
  6. Andreadis T, Gasteratos A, Seimenis I, Koulouriotis D. Application of CAD systems in breast cancer diagnosis using machine learning techniques: an overview of systematic reviews. Bioengineering. 2025;12(11):1160. doi:10.3390/bioengineering12111160 read 9 October 2026. doi.org/10.3390/bioengineering12111160
  7. Zhang J, Li H, Zhang Y, Huang J, Ren L, Zhang C, Zou Q, Zhang Y. Computational toxicology in drug discovery: applications of artificial intelligence in ADMET and toxicity prediction. Briefings in Bioinformatics. 2025;26(5):bbaf533. doi:10.1093/bib/bbaf533 read 9 October 2026. doi.org/10.1093/bib/bbaf533
  8. Fountzilas E, Pearce T, Baysal MA, Chakraborty A, et al. , Tsimberidou AM, et al. Convergence of evolving artificial intelligence and machine learning techniques in precision oncology. npj Digital Medicine. 2025;8(1):75. doi:10.1038/s41746-025-01471-y read 9 October 2026. doi.org/10.1038/s41746-025-01471-y
  9. Lee CF, Lin J, Huang YL, Chen ST, Chou CT, Chen DR, et al. , Wu WP, et al. Deep learning-based breast MRI for predicting axillary lymph node metastasis: a systematic review and meta-analysis. Cancer Imaging. 2025;25(1):44. doi:10.1186/s40644-025-00863-3 read 9 October 2026. doi.org/10.1186/s40644-025-00863-3
  10. Abdullah KA, Marziali S, Nanaa M, Escudero Sanchez L, Payne NR, et al. , Gilbert FJ, et al. Deep learning-based breast cancer diagnosis in breast MRI: systematic review and meta-analysis. European Radiology. 2025;35(8):4474-4489. doi:10.1007/s00330-025-11406-6 read 9 October 2026. doi.org/10.1007/s00330-025-11406-6
  11. Du S, Guo J, Huang D, Liu Y, Zhang X, Lu S. Diagnostic accuracy of deep learning-based algorithms in laryngoscopy: a systematic review and meta-analysis. European Archives of Oto-Rhino-Laryngology. 2025;282(1):351-360. doi:10.1007/s00405-024-09049-2 read 9 October 2026. doi.org/10.1007/s00405-024-09049-2
  12. Valizadeh P, Jannatdoust P, Pahlevan-Fallahy MT, Hassankhani A, Amoukhteh M, Bagherieh S, Ghadimi DJ, et al. , Gholamrezanezhad A, et al. Diagnostic accuracy of radiomics and artificial intelligence models in diagnosing lymph node metastasis in head and neck cancers: a systematic review and meta-analysis. Neuroradiology. 2025;67(2):449-467. doi:10.1007/s00234-024-03485-x read 9 October 2026. doi.org/10.1007/s00234-024-03485-x
  13. Yan F, Wu J, Li J, Wang W, Chen Y, Wei L, Lu J, Chen W, Gao Z, Li J, et al. , Wang Z, et al. PathOrchestra: a comprehensive foundation model for computational pathology with over 100 diverse clinical-grade tasks. npj Digital Medicine. 2025;8(1):695. doi:10.1038/s41746-025-02027-w read 9 October 2026. doi.org/10.1038/s41746-025-02027-w
  14. Escobar Diaz Guerrero R, Carvalho L, Bocklitz T, et al, et al. A data augmentation methodology to reduce the class imbalance in histopathology images. Journal of Imaging Informatics in Medicine. 2024. doi:10.1007/s10278-024-01018-9 read 9 October 2026. doi.org/10.1007/s10278-024-01018-9
  15. Vorontsov E, Bozkurt A, Casson A, Shaikovski G, Zelechowski M, Severson K, Zimmermann E, Hall J, Tenenholtz N, Fusi N, et al. , Fuchs TJ, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine. 2024;30(10):2924-2935. doi:10.1038/s41591-024-03141-0 read 9 October 2026. doi.org/10.1038/s41591-024-03141-0
  16. Xu H, Usuyama N, Bagga J, Zhang S, Rao R, Naumann T, Wong C, Gero Z, Gonzalez J, Gu Y, et al. , Poon H, et al. A whole-slide foundation model for digital pathology from real-world data. Nature. 2024;630(8015):181-188. doi:10.1038/s41586-024-07441-w read 9 October 2026. doi.org/10.1038/s41586-024-07441-w
  17. Catacutan DB, Alexander J, Arnold A, et al. , Stokes JM, et al. Machine learning in preclinical drug discovery. Nature Chemical Biology. 2024;20(8):960-973. doi:10.1038/s41589-024-01679-1 read 9 October 2026. doi.org/10.1038/s41589-024-01679-1
  18. Wang TW, Hong JS, Chiu HY, Chao HS, Chen YM, Wu YT. Standalone deep learning versus experts for diagnosis lung cancer on chest computed tomography: a systematic review. European Radiology. 2024;34(11):7397-7407. doi:10.1007/s00330-024-10804-6 read 9 October 2026. doi.org/10.1007/s00330-024-10804-6
  19. Sadybekov AV, Katritch V. Computational approaches streamlining drug discovery. Nature. 2023;616(7958):673-685. doi:10.1038/s41586-023-05905-z read 9 October 2026. doi.org/10.1038/s41586-023-05905-z
  20. Nasser M, Yusof UK. Deep learning based methods for breast cancer diagnosis: a systematic review and future direction. Diagnostics. 2023;13(1):161. doi:10.3390/diagnostics13010161 read 9 October 2026. doi.org/10.3390/diagnostics13010161
  21. Chou WC, Lin Z. Machine learning and artificial intelligence in physiologically based pharmacokinetic modeling. Toxicological Sciences. 2023;191(1):1-14. doi:10.1093/toxsci/kfac101 read 9 October 2026. doi.org/10.1093/toxsci/kfac101
  22. Tsimberidou AM, Kahle M, Vo HH, Baysal MA, Johnson A, et al. , Meric-Bernstam F, et al. Molecular tumour boards: current and future considerations for precision oncology. Nature Reviews Clinical Oncology. 2023;20(12):843-863. doi:10.1038/s41571-023-00824-4 read 9 October 2026. doi.org/10.1038/s41571-023-00824-4
  23. Bulten W, Kartasalo K, Chen PHC, Strom P, Pinckaers H, Nagpal K, Cai Y, Steiner DF, van Boven H, Vink R, Hulsbergen-van de Kaa C, van der Laak J, Amin MB, Evans AJ, van der Kwast T, Allan R, Humphrey PA, Gronberg H, Samaratunga H, Delahunt B, Tsuzuki T, Hakkinen T, Egevad L, Demkin M, Dane S, Tan F, Valkonen M, Corrado GS, Peng L, Mermel CH, Ruusuvuori P, Litjens G, Eklund M. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nature Medicine. 2022;28:154-63. read 9 October 2026. research.google
  24. Laleh NG, Muti HS, Loeffler CML, Echle A, Saldanha OL, Mahmood F, Lu MY, Trautwein C, Langer R, Dislich B, Buelow RD, Grabsch HI, Brenner H, Chang-Claude J, Alwers E, Brinker TJ, Khader F, Truhn D, Gaisa NT, Boor P, Hoffmeister M, Schulz V, Kather JN. Benchmarking weakly-supervised deep learning pipelines for whole slide classification in computational pathology. Medical Image Analysis. 2022;79. read 9 October 2026. arxiv.org
  25. Lipkova J, Chen TY, Lu MY, Chen RJ, Shady M, Williams M, Wang J, Noor Z, Mitchell RN, Turan M, Coskun G, Yilmaz F, Demir D, Nart D, Basak K, Turhan N, Ozkara S, Banz Y, Odening KE, Mahmood F. Deep learning-enabled assessment of cardiac allograft rejection from endomyocardial biopsies. Nature Medicine. 2022. read 9 October 2026. arxiv.org
  26. Ghahremani P, Li Y, Kaufman A, Vanguri R, Greenwald N, Angelo M, Hollmann TJ, Nadeem S. Deep learning-inferred multiplex immunofluorescence for immunohistochemical image quantification. Nature Machine Intelligence. 2022;4:401-12. read 9 October 2026. arxiv.org
  27. Huang SC, Chen CC, Lan J, Hsieh TY, Chuang HC, Chien MY, Ou TS, Chen KH, Wu RC, Liu YJ, Cheng CT, Huang YJ, Tao LW, Hwu AF, Lin IC, Hung SH, Yeh CY, Chen TC. Deep neural network trained on gigapixel images improves lymph node metastasis detection in clinical settings. Nature Communications. 2022;13. read 9 October 2026. arxiv.org
  28. Lee Y, Park JH, Oh S, Shin K, Sun J, Jung M, Lee C, Kim H, Chung JH, Moon KC, Kwon S. Derivation of prognostic contextual histopathological features from whole-slide images of tumours via graph deep learning. Nature Biomedical Engineering. 2022. read 9 October 2026. arxiv.org
  29. Lin Z, Chou WC. Machine learning and artificial intelligence in toxicological sciences. Toxicological Sciences. 2022;189(1):7-19. doi:10.1093/toxsci/kfac075 read 9 October 2026. doi.org/10.1093/toxsci/kfac075
  30. Chen RJ, Lu MY, Williamson DFK, Chen TY, Lipkova J, Noor Z, Shaban M, Shady M, Williams M, Joo B, Mahmood F. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell. 2022;40:865-78. read 9 October 2026. arxiv.org
  31. Krishnan R, Rajpurkar P, Topol EJ. Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering. 2022. read 9 October 2026. arxiv.org
  32. Pandey M, Fernandez M, Gentile F, Isayev O, Tropsha A, Stern AC, Cherkasov A. The transformational role of GPU computing and deep learning in drug discovery. Nature Machine Intelligence. 2022;4(3):211-221. doi:10.1038/s42256-022-00463-x read 9 October 2026. doi.org/10.1038/s42256-022-00463-x
  33. Wang X, Yang S, Zhang J, Wang M, Zhang J, Yang W, Huang J, Han X. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis. 2022;81. read 9 October 2026. arxiv.org
  34. Greenwald NF, Miller G, Moen E, Kong A, Kagel A, Dougherty T, Fullaway CC, McIntosh BJ, Leow KX, Schwartz MS, Pavelchek C, Cui S, Camplisson I, Bar-Tal O, Singh J, Fong M, Chaudhry G, Abraham Z, Moseley J, Warshawsky S, Soon E, Greenbaum S, Risom T, Hollmann T, Bendall SC, Keren L, Graf W, Angelo M, Van Valen D. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nature Biotechnology. 2022;40:555-65. read 9 October 2026. arxiv.org
  35. Lu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, Mahmood F. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594:106-10. read 9 October 2026. arxiv.org
  36. Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering. 2021;5:555-70. read 9 October 2026. arxiv.org
  37. Liu JTC, Glaser AK, Bera K, True LD, Reder NP, Eliceiri KW, Madabhushi A. Harnessing non-destructive 3D pathology. Nature Biomedical Engineering. 2021;5:203-18. read 9 October 2026. arxiv.org
  38. Wulczyn E, Steiner DF, Moran M, Plass M, Reihs R, Tan F, Flament-Auvigne I, Brown T, Regitnig P, Chen PHC, Hegde N, Sadhwani A, MacDonald R, Ayalew B, Corrado GS, Peng LH, Tse D, Muller H, Xu Z, Liu Y, Stumpe MC, Zatloukal K, Mermel CH. Interpretable survival prediction for colorectal cancer using deep learning. npj Digital Medicine. 2021;4. read 9 October 2026. arxiv.org
  39. Skrede OJ, De Raedt S, Kleppe A, Hveem TS, Liestol K, Maddison J, Askautrud HA, Pradhan M, Nesheim JA, Albregtsen F, Farstad IN, Domingo E, Church DN, Nesbakken A, Shepherd NA, Tomlinson I, Kerr R, Novelli M, Kerr DJ, Danielsen HE. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. The Lancet. 2020;395:350-60. read 9 October 2026. arxiv.org
  40. Wulczyn E, Steiner DF, Xu Z, Sadhwani A, Wang H, Flament-Auvigne I, Mermel CH, Chen PHC, Liu Y, Stumpe MC. Deep learning-based survival prediction for multiple cancer types using histopathology images. PLOS ONE. 2020;15:e0233678. read 9 October 2026. arxiv.org
  41. Lu Z, Xu S, Shao W, Wu Y, Zhang J, Han Z, Feng Q, Huang K. Deep-learning-based characterization of tumor-infiltrating lymphocytes in breast cancers from histopathology images and multiomics data. JCO Clinical Cancer Informatics. 2020;4:480-90. read 9 October 2026. arxiv.org
  42. Kather JN, Heij LR, Grabsch HI, Loeffler C, Echle A, Muti HS, Krause J, Niehues JM, Sommer KAJ, Bankhead P, Kooreman LFS. Pan-cancer image-based detection of clinically actionable genetic alterations. Nature Cancer. 2020;1:789-99. read 9 October 2026. arxiv.org
  43. Pinckaers H, van Ginneken B, Litjens G. Streaming convolutional neural networks for end-to-end learning with multi-megapixel images. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2020. read 9 October 2026. arxiv.org
  44. Liu X, Faes L, Kale AU, Wagner SK, Fu DJ, Bruynseels A, Mahendiran T, Moraes G, Shamdas M, Kern C, Ledsam JR, Schmid MK, Balaskas K, Topol EJ, Bachmann LM, Keane PA, Denniston AK. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. The Lancet Digital Health. 2019;1(6):e271-e297. doi:10.1016/S2589-7500(19)30123-2 read 9 October 2026. doi.org/10.1016/S2589-7500(19)30123-2
  45. Campanella G, Hanna MG, Geneslaw L, Miraflor A, Werneck Krauss Silva V, Busam KJ, Brogi E, Reuter VE, Klimstra DS, Fuchs TJ. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine. 2019. read 9 October 2026. arxiv.org
  46. Abels E, Pantanowitz L, Aeffner F, Zarella MD, van der Laak J, Bui MM, Vemuri VNP, Parwani AV, Gibbs J, Agosto-Arroyo E, Beck AH, Kozlowski C. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the Digital Pathology Association. The Journal of Pathology. 2019;249:286-94. read 9 October 2026. arxiv.org
  47. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine. 2019;25(1):44-56. doi:10.1038/s41591-018-0300-7 read 9 October 2026. doi.org/10.1038/s41591-018-0300-7
  48. Graham S, Vu QD, Raza SEA, Azam A, Tsang YW, Kwak JT, Rajpoot N. HoVer-Net: simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis. 2019;58. read 9 October 2026. arxiv.org
  49. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Medicine. 2019. doi:10.1186/s12916-019-1426-2 read 9 October 2026. doi.org/10.1186/s12916-019-1426-2
  50. Veta M, Heng YJ, Stathonikos N, Ehteshami Bejnordi B, Beca F, Wollmann T, Rohr K, Shah MA, Wang D, Rousson M, Hedlund M, Tellez D, Ciompi F, Zerhouni E, Lanyi D, Viana M, Kovalev V, Liauchuk V, Phoulady HA, Qaiser T, Graham S, Rajpoot N, Sjoblom E, Molin J, Paeng K, Hwang S, Park S, Jia Z, Chang EIC, Xu Y, Beck AH, van Diest PJ, Pluim JPW. Predicting breast tumor proliferation from whole-slide images: the TUPAC16 challenge. Medical Image Analysis. 2019;54:111-21. read 9 October 2026. arxiv.org
  51. Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyo D, Moreira AL, Razavian N, Tsirigos A. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nature Medicine. 2018;24(10):1559-67. doi:10.1038/s41591-018-0177-5 read 9 October 2026. doi.org/10.1038/s41591-018-0177-5
  52. Bandi P, Geessink O, Manson Q, van Dijk M, Balkenhol M, Hermsen M, Ehteshami Bejnordi B, Lee B, Paeng K, Zhong A, Li Q. From detection of individual metastases to classification of lymph node status at the patient level: the CAMELYON17 challenge. IEEE Transactions on Medical Imaging. 2018;38:550-60. read 9 October 2026. arxiv.org
  53. Saltz J, Gupta R, Hou L, Kurc T, Singh P, Nguyen V, Samaras D, Shroyer KR, Zhao T, Batiste R, Van Arnam J. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Reports. 2018;23:181-93. read 9 October 2026. arxiv.org
  54. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115-118. doi:10.1038/nature21056 read 9 October 2026. doi.org/10.1038/nature21056
  55. Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, van der Laak JAWM, and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA. 2017;318:2199-210. read 9 October 2026. arxiv.org
  56. Sirinukunwattana K, Pluim JPW, Chen H, Qi X, Heng PA, Guo YB, Wang LY, Matuszewski BJ, Bruni E, Sanchez U, Bohm A, Ronneberger O, Cheikh BB, Racoceanu D, Kainz P, Pfeiffer M, Urschler M, Snead DRJ, Rajpoot NM. Gland segmentation in colon histology images: the GlaS challenge contest. Medical Image Analysis. 2017;35:489-502. read 9 October 2026. arxiv.org
  57. Madabhushi A, Lee G. Image analysis and machine learning in digital pathology: challenges and opportunities. Medical Image Analysis. 2016;33:170-5. read 9 October 2026. arxiv.org
  58. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436-44. read 9 October 2026. arxiv.org
  59. Beck AH, Sangoi AR, Leung S, Marinelli RJ, Nielsen TO, van de Vijver MJ, West RB, van de Rijn M, Koller D. Systematic analysis of breast cancer morphology uncovers stromal features associated with survival. Science Translational Medicine. 2011;3. read 9 October 2026. arxiv.org

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 9). Machine Learning in Computational Pathology for the Pharmaceutical Industry. Pubrica. https://pubrica.com/academy/uncategorized/how-is-machine-learning-significant-to-computational-pathology-in-the-pharmaceutical-industries/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Computational pathology study record

This record should be completed before the methods are written. Each line is included because a reviewer, a purchaser or a regulator will ask the question, and because a study that cannot answer it cannot be reproduced by others.

COMPUTATIONAL PATHOLOGY STUDY RECORD

1. THE CLAIM  (pick ONE - they need different reference standards)
   [ ] Agreement with a reference read (grading, detection)
         Reference standard: ______________________
   [ ] Inference of a molecular property from morphology
         Orthogonal assay used as standard: ______________
         Assay error rate, if known: ______
   [ ] Prediction of outcome from morphology
         Cohort, follow-up duration, events: ______________
         [ ] Incremental value over stage/grade demonstrated
   Intended use, stated at the level of specimen type, stain,
   scanner and clinical question:
   ______________________________________________________

2. SPECIMEN CHAIN  (fill in for EVERY data source)
   Source A (development):
     Cases ______  Fixation ____________  Stain platform ________
     Protocol ____________  Section thickness ______
     Scanner and model ____________  Magnification ______
     Compression ______
   Source B (external validation):
     Cases ______  Fixation ____________  Stain platform ________
     Protocol ____________  Section thickness ______
     Scanner and model ____________  Magnification ______
     [ ] Source B contributed NOTHING to model development

3. PARTITIONING
   [ ] Patient-level separation between all partitions
   [ ] Held-out set from an INDEPENDENT LABORATORY
   [ ] Held-out set from a later time period
   Development: ____________  Tuning: ____________
   External test set(s): ____________

4. MODEL AND AGGREGATION
   Supervision: [ ] fully supervised  [ ] weakly supervised
                [ ] self-supervised pretraining  [ ] foundation model
   Foundation model used, and version: ______________________
   Tile size and magnification: ______________________
   Aggregation from tile to slide: ______________________
   Aggregation from slide to patient: ______________________
   [ ] Aggregation function fixed BEFORE test evaluation
   Class imbalance handling: ______________________
   [ ] Effect of imbalance handling on output probabilities stated

5. RESULTS  (per test set, per stain platform, per scanner)
   Operating point ______  Fixed before test evaluation? [ ]
                              Internal      External
   AUC ......................  ________      ________
   Sensitivity ..............  ________      ________
   Specificity ..............  ________      ________
   Calibration slope ........  ________      ________
   Agreement statistic ......  ________      ________
   [ ] Transfer gap reported explicitly, not averaged away
   [ ] Results NOT pooled across laboratories
   Prevalence per set: ______    Intended-use prevalence: ______

6. REGULATORY POSTURE  (if intended for use)
   Model version locked?  [ ] yes, version ______
   Change control for retraining: ______________________
   Reporting standard: [ ] TRIPOD+AI  [ ] STARD-AI
                       [ ] CONSORT-AI / SPIRIT-AI
                       [ ] DECIDE-AI  [ ] PROBAST+AI
   [ ] Digital Pathology Association best-practice items addressed

7. LIMITATIONS  (one specific sentence each)
   Preanalytical variation: ______________________
   Agreement vs clinical benefit: ______________________
   Operating point and prevalence: ______________________
   Task scope of any foundation model used: ______________

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s pathology and data teams address the aspects of a computational pathology study that determine whether it transfers: full description of the specimen chain, validation on material from a laboratory that contributed nothing to development, a fixed and reported aggregation function and operating point, and performance reported by stain platform and scanner rather than pooled. Pubrica also separates the three claims (agreement, molecular inference and prognosis) so that each is validated against the appropriate standard, and maps the study to the reporting guideline that a regulator or reviewer will apply.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

On receipt of a pathology model and validation plan, Pubrica assesses whether the result is likely to hold at another laboratory.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.