Computational pathology overview

Computational pathology uses digitized tissue and associated clinical or molecular data to study morphology, diagnosis, biomarkers, prognosis, and clinical workflows. Its defining technical object is often the whole-slide image (WSI): a multiresolution acquisition that can contain billions of pixels and many biologically distinct regions.

Latest literature discovery pass: 12 October 2026, across all ten pathology routes. Selected additions received full source reviews; existing papers retain their individual source-check dates. This is not a re-verification of every paper or an exhaustive publication census.

This section is a maintained research route, not one model leaderboard. Each chapter combines concepts, dated primary papers, fair-comparison rules, and a reproducible project. Use the curated paper guide for the cross-field route and My Library to import a category into your private tracker.

The searchable literature registry filters the canonical computational-pathology records by year, evidence state, and research route. It is also the source used by My Library, so the public catalog and personal picker stay in sync.

The field in one pipeline

patient → case → specimen → block → slide → WSI pyramid
                                              │
                    ┌─────────────────────────┼──────────────────────────┐
                    ↓                         ↓                          ↓
             cells / tissue             tile encoders             image + text/omics
                    │                         │                          │
                    └──────────────→ slide / spatial model ←────────────┘
                                              │
                                              ↓
                     diagnosis · grade · biomarker · outcome · workflow
                                              │
                                              ↓
                       external evidence · readers · prospective use

Every arrow can introduce leakage, measurement error, shortcut learning, missingness, or a mismatch between the claimed and actual population.

The core learning route

Stage	Dedicated chapter	Central question	Readiness output
0. Pathology and data	this overview + WSI Data Atlas	what physical specimen, stain, hierarchy, and label created the data?	cohort diagram and data dictionary
1. WSI systems	WSI engineering and QC	how do acquisitions become reproducible tiles and embeddings?	versioned slide/QC/tile manifest
2. Dense morphology	cell, gland, and tissue modeling	what requires pixels, points, instances, types, or compartments?	annotation and object-metric protocol
3. Weak slide labels	MIL and whole-slide learning	how is sparse, heterogeneous evidence aggregated?	controlled frozen-feature MIL comparison
4. Representation objectives	pathology self-supervised learning	what invariance, prediction, or hierarchy creates useful features?	objective and augmentation audit
5. Reusable models	foundation encoders	which exact tile/slide checkpoint fits the task and access constraints?	release-pinned model card and benchmark
6. Language and generation	pathology VLMs	how are image, report, caption, question, and slide evidence aligned?	grounding and leakage evaluation
7. Biology and outcomes	spatial, molecular, and outcome modeling	how can morphology relate to omics, spatial systems, and censored outcomes?	claim-matched multimodal/spatial study
8. Robust deployment	robustness, domain shift, and privacy	how are site/scanner shift, monitoring, federation, and privacy claims tested?	external-shift, monitoring, and threat-model protocol
9. Translation	evaluation and clinical translation	what evidence supports the intended use?	external, calibrated, fully reported evaluation

The stages are not a chronological history. A cell-segmentation project may branch at stage 2; a slide foundation-model project needs stages 1, 3, 4, 5, 8, and 9; a report generation project also needs stage 6.

1. Learn enough pathology to define the target

Understand:

• specimen preparation, fixation, sectioning, H&E, special stains, and IHC;
• cell, gland, tissue-compartment, slide, case, and patient hierarchy;
• grade versus stage;
• diagnostic, prognostic, and predictive targets;
• resection, biopsy, frozen section, cytology, and tissue microarray differences; and
• which information a pathologist has at the intended prediction time.

A label extracted from a report may represent the case, while the relevant morphology appears on only one block or slide. A molecular result may be measured on another section or at another time. Treat label construction as a measurement study.

Public background starting points include PathOER, the CAP Digital Pathology Topic Center, and ESDIP Academy. They support orientation; they do not replace pathology training or local expert collaboration.

2. Choose the unit and supervision

Available supervision	Typical task	Suitable starting family	Key failure
pixel/instance labels	tumour, gland, nucleus, cell type	dense segmentation/detection	annotation and domain shift
region labels	tissue class, lesion, grade	supervised patch/region model	spatial selection bias
slide/case labels	diagnosis, biomarker, outcome	MIL or slide-context model	sparse evidence and leakage
unlabelled images	transferable morphology	SSL/foundation encoder	harmful invariance or overlap
paired image–text	retrieval, zero-shot class, caption, QA	dual encoder or generative VLM	noisy text and ungrounded answers
paired image–omics	molecular/spatial inference or fusion	multimodal/graph model	assay alignment and confounding
time-to-event	prognosis	survival MIL/graph/fusion	censoring, treatment, calibration

Start with the weakest model that matches the supervision and scientific question. Keep encoder, sampling, and aggregator comparisons separate where possible.

3. Treat WSI processing as part of the model

Record scanner metadata, MPP, stain, pyramid/read level, tissue mask, tile dimensions, coordinates, sampling, QC decisions, and patient linkage. Core tools and standards:

• OpenSlide for common vendor formats;
• DICOM WSI for interoperable representation;
• TIAToolbox for computational workflows;
• HistoQC and GrandQC for QC;
• QuPath for viewing, annotation, and analysis.

The WSI engineering chapter turns this list into an auditable pipeline. Split patients before tiling.

4. Representation and aggregation are different questions

A complete slide system often has:

tile selection → tile encoder → optional contextualizer → aggregator → task head

An improvement can come from any component. Compare:

• exact encoder release and preprocessing;
• tile scale, field of view, and sampling;
• frozen, adapter, or fine-tuned transfer;
• aggregator and context budget;
• patient/site/temporal split;
• pretraining overlap;
• compute, storage, latency, and failure rate.

Attention weights can aid inspection, but they are not automatically calibrated lesion localization, faithful explanation, or causal attribution.

5. Keep claims at the evidence level

A retrospective internal test does not establish clinical utility. Progress through external and temporal data, calibration and operating points, reader/workflow studies, prospective evaluation, and monitoring as the intended use requires.

High-level questions:

• Does the model learn morphology or scanner/site signatures?
• Does color handling improve transport or only the internal test?
• Are gains robust to another encoder, aggregator, and seed?
• Was the endpoint available at prediction time?
• Are related patients, slides, serial sections, or public pretraining data shared?
• Do performance and calibration hold by site, specimen, and clinically relevant group?
• What happens to a missing, corrupt, low-quality, or low-confidence case?
• What user action changes, and what is the consequence of an error?

Minimum experiment record

intended_use: user, population, input, output, action
population: eligibility, sites, dates, exclusions
target: source, assay/annotator, timing, unit
hierarchy: patient > case > specimen > block > slide > tile/object
split_unit: patient_or_higher
external_test: site, time, specimen and overlap audit
wsi_pipeline: reader, QC, MPP, mask, sampling, stain policy
model: exact checkpoint, license, preprocessing
adaptation: frozen | linear | adapter | fine_tuned
aggregator: exact method and context budget
uncertainty: confidence interval and resampling unit
calibration: method, fit set, operating point
failures: exclusions, abstentions, fallback

Suggested first month

1. Read this overview and make a one-page specimen/label diagram for your project.
2. Select one dataset family in the WSI Data Atlas and verify its official access, hierarchy, license, and warnings.
3. Complete the WSI engineering manifest exercise.
4. Follow either the dense-morphology or MIL route based on the available label.
5. Read the SSL and encoder chapters before selecting a checkpoint.
6. Write the intended-use and evidence table from the evaluation chapter before reporting performance.
7. Use the robustness, shift, and privacy chapter before making a cross-site, federated, monitoring, or confidentiality claim.

The goal is not to accumulate model names. It is to connect a biological and clinical question to a traceable specimen, appropriate supervision, reproducible representation, fair comparison, and evidence strong enough for the stated claim.

Generative and virtual pathology

Synthetic patches, cell layouts, gigapixel slides, virtual stains, and cross-modal molecular generation require a different validation contract from discriminative prediction. Continue to generative and virtual pathology for multiscale coherence, morphology and topology controls, privacy tests, pathologist review, and downstream utility on independent real patients.

Robustness, shift, and privacy

Site, scanner, stain, population, time, and workflow changes require more than one average external score. Federated data locality also does not by itself establish privacy. Continue to robustness, domain shift, and privacy for nuisance interventions, natural-shift benchmarks, model-aware monitoring, federated WSI systems, explicit privacy boundaries, and a deployment response loop.