Global Pathology Review — Path-iQ · Reference Guide
Everything you need to understand artificial intelligence in pathology: from whole-slide imaging infrastructure and foundation model architectures to FDA-cleared tools, clinical deployment strategies, and the future of computational histopathology.
Definition
AI pathology — also called computational pathology or CPath — is the application of artificial intelligence, machine learning, and computer vision to the analysis of histopathological tissue images. It represents one of the most significant technological shifts in diagnostic medicine since the introduction of immunohistochemistry in the 1970s.
At its core, AI pathology involves training deep neural networks on large collections of digitised glass slides to recognise and quantify morphological patterns that pathologists have traditionally identified by eye under the microscope. These patterns include tumour presence, grade, subtype, and a growing range of predictive molecular biomarkers that can be inferred directly from haematoxylin and eosin (H&E) stained tissue without any additional molecular testing.
The field builds on decades of pathology informatics research but underwent a step-change after 2016 when deep convolutional neural networks — the same architectures that transformed image recognition in natural photography — were applied to histopathological images with remarkable accuracy. Landmark papers from groups at Google, Stanford, Case Western Reserve University, and multiple European academic centres demonstrated that AI systems could perform prostate cancer Gleason grading, breast cancer HER2 scoring, and lymph node metastasis detection at or above the accuracy of expert pathologists.
Key distinction: AI pathology is not a replacement for pathologists. It is most accurately understood as a precision instrument — capable of exhaustive, quantitative, reproducible analysis of every cell in a slide — that augments human diagnostic expertise and frees pathologists for higher-order clinical reasoning.
Identifying and outlining specific tissue regions, cell types, or lesions. Examples: tumour detection in lymph nodes, gland segmentation in prostate biopsies, nucleus counting in Ki-67 IHC.
Assigning diagnostic categories or severity scores to tissue. Examples: Gleason grading, ISUP grade grouping, WHO CNS tumour subtype classification, breast cancer ER/PR H-score quantification.
Inferring molecular or genomic features directly from H&E morphology. Examples: MSI status, TMB, BRAF V600E, HER2 amplification, and mutational signature detection — without additional molecular testing.
Definition
Digital pathology is the practice of digitising conventional glass microscopy slides into high-resolution digital images called whole-slide images (WSIs), then managing, distributing, viewing, and analysing those images on digital platforms rather than — or in addition to — optical microscopes.
A WSI is an extraordinarily large image: a single 20× scan of a standard histology section typically produces a file of 1–3 gigabytes, containing hundreds of millions of pixels at a resolution sufficient to resolve individual cell nuclei and their chromatin patterns. At 40×, the same slide generates 4–10 GB. The management, storage, and efficient delivery of WSIs at scale is a substantial informatics challenge that has given rise to a specialised ecosystem of image servers, viewers, LIMS integrations, and cloud infrastructure.
Digital pathology is the infrastructure layer — it enables slides to exist as digital files. AI pathology is one of several things you can do with those files. Other applications of digital pathology include remote primary diagnosis (telepathology), multidisciplinary tumour board presentations on screens, quantitative image analysis without AI, annotation tools for research, and digital archiving. AI pathology requires digital pathology as its foundation, but digital pathology itself does not require AI.
"Digital pathology is to AI pathology what the internet was to e-commerce — necessary infrastructure, not the application itself."
— Path-iQ Editorial, 2026
| Year | Milestone |
|---|---|
| 2000 | First commercial whole-slide scanners introduced (Aperio/Hamamatsu) |
| 2009 | FDA begins evaluating digital pathology for primary diagnosis |
| 2017 | FDA clears first digital pathology system for primary diagnosis (Philips IntelliSite) |
| 2019 | Paige.AI receives breakthrough device designation from FDA |
| 2021 | Paige Prostate becomes first AI pathology tool with FDA De Novo clearance |
| 2022 | Foundation model era begins: PLIP, CONCH, UNI published on arXiv |
| 2023 | TITAN and CTransPath multimodal models published; FDA clears additional AI tools |
| 2024 | AI pathology market surpasses $800M; major health systems complete full digital transitions |
| 2025 | WHO classification guidelines reference AI-assisted diagnosis for the first time |
| 2026 | Estimated 35% of pathology laboratories globally operating some form of digital workflow |
Technical primer
Modern AI pathology is dominated by deep learning — specifically convolutional neural networks (CNNs) and, increasingly, vision transformer (ViT) architectures. The core technical challenge is unique to pathology: unlike natural photographs, WSIs are gigapixel images far too large to process as a single input. The field has converged on several strategies for handling this scale.
The most common approach. The WSI is divided into thousands of small image patches (typically 256×256 or 512×512 pixels), each patch is independently classified by a CNN or ViT, and results are aggregated across the slide. This approach is computationally efficient and works well for spatially localised tasks like gland grading or metastasis detection.
For slide-level classification tasks (e.g., "does this slide contain cancer?"), where patch-level labels are unavailable, MIL treats each slide as a "bag" of patches with a single bag-level label. The model learns which patches are most informative via attention mechanisms. The ABMIL and TransMIL architectures are widely used MIL frameworks in pathology.
ViTs divide the image into fixed-size patches treated as tokens (analogous to words in NLP) and apply self-attention to model relationships between distant image regions. This global context awareness makes ViTs well-suited for understanding the spatial architecture of tumours — a property that CNNs, which are local, lack. DINO-v2 and H-optimus-0 are widely used ViT backbones in 2025–2026 pathology models.
AI pathology models require large, expertly annotated training datasets. The largest public dataset is The Cancer Genome Atlas (TCGA), containing ~11,000 WSIs from 33 cancer types with corresponding molecular data. Proprietary datasets at large health systems (MSKCC, MGH, Karolinska) are substantially larger. Training annotation is the primary bottleneck: a single prostate grading dataset may require 10,000+ hours of pathologist time to create.
Histological slides vary substantially in colour due to differences in tissue processing, staining reagent brands, scanner hardware, and laboratory protocols. This stain variability, if unaddressed, severely degrades AI model performance when deployed across institutions. Stain normalisation algorithms (Macenko, Vahadane, cycle-GAN-based approaches) or stain-augmentation during training are standard preprocessing steps.
State of the art
A foundation model in pathology is a large neural network pre-trained on vast quantities of histopathological images in a self-supervised manner — without task-specific labels — that can subsequently be adapted (fine-tuned) to a wide range of downstream diagnostic tasks with relatively little labelled data. The term, borrowed from the NLP field where GPT and BERT pioneered the paradigm, has transformed how pathology AI models are built since 2022.
Rather than training a new model from scratch for each diagnostic task (which requires thousands of labelled examples), a foundation model provides a rich visual representation of tissue morphology that transfers broadly. A lab wanting to build an endometrial carcinoma grading tool can fine-tune a pathology foundation model with hundreds of labelled slides rather than thousands.
| Model | Organisation | Architecture | Training Data | Modality | Year |
|---|---|---|---|---|---|
| UNI | Harvard / MGH | ViT-Large (DINOv2) | 100k+ WSIs, 20+ tissue types | Vision | 2024 |
| TITAN | UBC / Mahmood Lab | ViT + LLM | 335k WSIs + reports (TCGA scale) | Multimodal | 2024 |
| CONCH | Harvard / MGH | ViT-Base (CLIP) | 1.17M image-text pairs | Vision-Language | 2024 |
| PLIP | Emory | ViT-Base (CLIP) | 208k Twitter image-caption pairs | Vision-Language | 2023 |
| CTransPath | HKUST | Swin Transformer | 15M patches from 32k WSIs | Vision | 2022 |
| H-optimus-0 | Owkin | ViT-g (DINOv2) | 500k+ WSIs | Vision | 2024 |
| Prov-GigaPath | Microsoft / Providence | ViT (LongNet) | 170k WSIs, 1B+ patches | Vision | 2024 |
| PathChat | Harvard / MGH | LLaVA-style | UNI features + pathology Q&A | Vision-Language Chat | 2024 |
The TITAN model (Nature Medicine, 2024) from the Mahmood Lab at Harvard demonstrated that training on whole-slide images paired with pathology reports — using a multimodal contrastive objective — produces features that outperform patch-level pre-training across nearly all downstream benchmarks, representing the current state of the art in generalised pathology representation learning.
Regulatory landscape
AI pathology tools intended for clinical use in the United States require FDA marketing authorisation as Software as a Medical Device (SaMD). The most common pathways are 510(k) clearance (substantial equivalence to a predicate device) and De Novo classification (for novel device types without a predicate). Tools used only for research or workflow efficiency, not for diagnosis, may not require clearance.
| Product | Company | Indication | Pathway | Year |
|---|---|---|---|---|
| Paige Prostate | Paige.AI | Prostate adenocarcinoma detection | De Novo | 2021 |
| Paige Breast | Paige.AI | Breast carcinoma detection | 510(k) | 2023 |
| Paige Colorectal | Paige.AI | Colorectal carcinoma detection | 510(k) | 2023 |
| Ibex Prostate | Ibex Medical Analytics | Prostate cancer Gleason grading | CE-IVD (EU) / 510(k) pending | 2022 |
| Proscia Concentriq AI | Proscia | Prostate, breast workflow AI | 510(k) | 2023 |
| PathAI AISight | PathAI | NASH/NAFLD grading | Breakthrough device | 2024 |
| Halo AP | Indica Labs | IHC quantification (PD-L1, Ki-67) | LDT / CE-IVD | 2022 |
Table reflects publicly announced regulatory status as of July 2026. CE-IVD = EU In Vitro Diagnostic Regulation compliance. LDT = laboratory-developed test. Regulatory landscape changes frequently; verify current status with manufacturers.
Applications
The most mature AI pathology application. Multiple validated tools perform Gleason grading of prostate needle biopsies at expert pathologist-level accuracy. The landmark 2019 study from Nagpal et al. (Google/UCSF, JAMA Oncology) demonstrated that AI-assisted grading outperformed general pathologists and matched specialist uropathologist accuracy. Clinical deployment at systems including Johns Hopkins, Karolinska, and Cleveland Clinic is now routine. AI tools are particularly valued for detecting small foci of cancer in long biopsy cores that may be missed on rapid screening, and for ensuring Gleason grade group consistency across high-volume laboratories.
AI applications in breast pathology span tumour detection, histological subtype classification, mitotic figure counting, Ki-67 quantification, TIL (tumour-infiltrating lymphocyte) scoring, and HER2 IHC scoring. TIL assessment is particularly important: manual TIL scoring is subjective and reader-variable, while AI tools produce consistent, reproducible TIL percentages that are directly incorporated into clinical trial endpoints (KEYNOTE, IMpassion series). AI-assisted HER2 IHC scoring improves concordance between 2+ equivocal cases sent for ISH reflex testing.
NSCLC subtype classification (adenocarcinoma vs. squamous cell carcinoma vs. others) from H&E slides is an active AI pathology application, particularly relevant for small biopsy specimens where tissue is too limited for a full IHC panel. Beyond morphological subtyping, AI tools can predict EGFR, ALK, KRAS, and STK11 mutation status directly from H&E slides with clinically meaningful accuracy, opening the possibility of AI-assisted triage of molecular testing requests.
AI models for DLBCL cell-of-origin (COO) prediction from H&E, germinal centre B-cell vs. activated B-cell classification, and automated CD30/MUM1/BCL6 IHC scoring are in active development. The 2024 Saldanha et al. (Lancet Digital Health) study demonstrated AI prediction of DLBCL molecular subtypes from H&E slides, a finding with significant implications for low-resource settings where GEP testing is unavailable.
NASH activity scoring (NAS) and fibrosis staging (Metavir/Ishak/Kleiner) are notoriously variable between pathologists, creating reproducibility problems in clinical trials. PathAI's NASH AI received FDA breakthrough device designation for automated NAS scoring, reducing inter-reader variability in drug trials from ±2 to ±0.4 grade points in prospective validation studies.
Perhaps the most transformative long-term application. Studies published in Nature Medicine and Lancet Oncology from 2019–2024 have demonstrated that AI models can predict MSI (microsatellite instability), TMB (tumour mutation burden), homologous recombination deficiency (HRD), BRAF V600E, and specific mutational signatures — all from standard H&E slides. If validated prospectively, these "virtual biomarkers" could provide a rapid, low-cost prescreening layer before expensive molecular testing, particularly valuable in resource-limited global settings.
Multiple FDA-cleared tools; clinical deployment at major academic centres worldwide.
Used in KEYNOTE and IMpassion trial endpoints; reproducible AI scoring adopted by IBEX and Paige.
Multiple validated models (STAMP-2, Kather et al.); awaiting prospective clinical validation.
H&E-based ADC vs. SCC classification validated on small biopsies; clinical implementation at select centres.
WSI-based survival models outperform clinical staging in multiple cancer types (TCGA benchmarks).
AI models predicting gene expression patterns from H&E morphology at spot resolution.
Infrastructure
Implementing digital pathology requires transforming a traditional glass-slide-and-microscope laboratory into one where slides are digitised and reviewed on screens. The workflow involves six core steps:
Conventional FFPE (formalin-fixed paraffin-embedded) tissue processing, sectioning, H&E staining, and glass slide preparation. No changes are required here — digital pathology works with standard glass slides.
Slide scanners (Aperio GT450, Hamamatsu NanoZoomer, Philips IntelliSite, Leica Aperio, 3DHistech P1000) digitise slides at 20× or 40× magnification. Throughput ranges from 60 slides/hour (standard) to 600 slides/hour (high-throughput multi-slide scanners). Scanning time per slide: 60–180 seconds at 20×.
A Digital Management System (or Image Management System) receives, stores, organises, and tracks WSI files. Integration with the laboratory information system (LIS) links images to case metadata. Storage requirements are substantial: a 500-bed hospital generating 30,000 slides/year requires ~100 TB of WSI storage per year. Cloud storage (AWS, Azure, Google Cloud with specialised pathology layers) or on-premises SAN solutions are both used.
Pathologists review WSIs on calibrated high-resolution monitors using specialised viewers (Sectra, Philips Pathologiq, Proscia Concentriq, Aperio ePathology). Ergonomics of digital review — including monitor size, brightness, zoom navigation, and screen calibration — have been studied extensively. ISO 15004 and DICOM standards apply to image capture and display.
AI algorithms are applied to WSIs either at scan time (pre-analysis), on demand (when a case is opened), or batch-processed overnight. Results are overlaid on the slide as heat maps, annotation layers, or quantitative measurements. Integration with the LIS allows AI results to pre-populate pathology report templates.
The pathologist reviews AI outputs, accepts or overrides them, dictates or types the final report, and electronically signs out. All interactions are logged for quality audit. AI-assisted report turnaround time (TAT) improvements of 15–30% have been reported in prospective studies.
Market landscape
The digital and AI pathology market encompasses scanner manufacturers, image management platform providers, AI algorithm developers, and integrated platform companies. The landscape has consolidated significantly since 2020 through acquisitions (Leica/Aperio, Hologic/Digital Scientific, Roper Technologies/Indica Labs).
| Company | Category | Key Products | Notes |
|---|---|---|---|
| Leica Biosystems | Scanner + LIS | Aperio GT450, Aperio ePathology, LIS integration | Market leader in slide scanning volume |
| Hamamatsu | Scanner | NanoZoomer S360, S60MD (FDA cleared) | Strong in academic research and Japan |
| Philips | Scanner + Platform | IntelliSite Pathologist Suite | First FDA-cleared digital pathology system (2017) |
| 3DHistech | Scanner + Viewer | P1000, PANNORAMIC series, CaseViewer | High throughput, strong European presence |
| Paige.AI | AI algorithms | Paige Prostate, Breast, Colorectal | First FDA De Novo cleared AI pathology tool |
| Ibex Medical | AI algorithms | Galen Prostate, Galen Breast | CE-IVD; clinical use across EU and Middle East |
| PathAI | AI + CRO services | AISight NASH, TIL scoring, clinical trials | Major pharma partnerships (BMS, J&J, GSK) |
| Proscia | Platform + AI | Concentriq AP, Concentriq AI | Enterprise digital pathology management |
| Indica Labs | Quantitative IHC | HALO, HALO AP | Widely used research and clinical IHC quantification |
| Sectra | Viewer + PACS | Sectra Digital Pathology | Integrated radiology + pathology PACS |
| Tribun Health | AI + Platform | Cytomine, SupraPathology | CE-IVD; strong European academic partnerships |
| Aiforia | AI platform | Aiforia Create, Deploy | No-code AI model training for pathologists |
Barriers & limitations
Each AI algorithm requires independent analytical and clinical validation before clinical deployment. The FDA expects prospective, multi-site validation data with defined performance endpoints and failure mode analysis. For most diagnostic applications, this means thousands of cases across multiple laboratory protocols — a multi-million dollar investment before any clinical revenue is generated.
AI models trained on data from one scanner, staining protocol, or patient population often perform poorly when deployed on data from a different laboratory environment. This "distribution shift" problem is the most significant practical barrier to broad AI pathology adoption. Organisations like the Digital Pathology Association (DPA) have published stain normalisation and image quality benchmarks to address this.
Creating ground-truth labelled datasets requires expert pathologist time — the most constrained and expensive resource in the field. Rare tumour subtypes may have insufficient training data for reliable AI model development. Federated learning (training models across multiple institutions without sharing patient data) and synthetic data generation are active research areas addressing this bottleneck.
Even validated AI tools face adoption barriers. Pathologists report concerns about liability when AI outputs differ from their own assessment, cognitive friction from poorly designed user interfaces, and lack of explainability in AI outputs. AI tools that explain their reasoning through attention maps or case-similar retrieval ("this was classified as Gleason 4 because of the following gland morphology features") show substantially better adoption rates than black-box outputs.
A full digital pathology implementation at a 500-slide-per-day laboratory requires investment in slide scanners ($200k–$500k each), image storage ($50k–$200k/year), LIS integration, high-resolution monitors, and network bandwidth. Total implementation costs at a major academic centre typically range from $3M–$10M. Cloud-based deployment models are lowering barriers but introduce data sovereignty and latency concerns.
The label problem: The entire field is bottlenecked by the need for expert pathologist annotations. A single prostate grading dataset sufficient for model training may require 50,000+ individual core-level labels, representing 2–3 years of part-time pathologist annotation work. Self-supervised learning (the foundation model approach) is the most promising technical solution — it learns from unlabelled data, reserving expert annotation for final calibration.
Outlook
The next frontier is models that jointly process histopathological images, genomic sequencing data, radiology images, and clinical records to produce integrated diagnostic and prognostic outputs. The TITAN and MUSK models represent early examples. Within five years, the expectation is that AI will routinely generate "integrated molecular-morphological" reports that currently require weeks of laboratory work, in minutes from a single H&E slide.
Large language models with vision capability (PathChat, GPT-4V applied to pathology) are being developed to draft full diagnostic reports from WSI inputs. Early studies show LLM-generated reports capture the correct diagnosis in 80–90% of cases, with pathologist editing reducing error to clinical standard. AI-assisted report generation could reduce pathologist dictation time by 50–70% while simultaneously improving structured data capture for downstream analytics.
Spatial transcriptomics (10x Visium, MERFISH, CosMx) generates gene expression data at cellular resolution within tissue sections. AI models that learn to predict spatial gene expression patterns from H&E morphology — creating "spatial virtual stains" — represent a particularly exciting research direction. Early publications (He et al., Nature Methods 2020; Schmauch et al., Nature Communications 2020) have demonstrated proof of principle. Clinical deployment at scale remains 5–10 years away.
The global shortage of pathologists is most acute in sub-Saharan Africa, South Asia, and parts of Latin America, where pathologist-to-population ratios can be 1:1,000,000 or worse. Tablet-based whole-slide viewers, low-cost smartphone microscope adaptors, and foundation models that work across diverse tissue processing protocols are enabling AI pathology deployment in settings where the traditional digital pathology infrastructure is impossible. Initiatives including Path AI for Africa and the Global Pathology Alliance are piloting these approaches.
The FDA's SaMD action plan and the EU AI Act (effective 2026) are creating structured regulatory frameworks for AI pathology tools. The trend is toward post-market performance monitoring requirements ("locked vs. adaptive algorithms"), manufacturer real-world performance reporting, and increasing emphasis on algorithmic transparency. The CAP has published AI pathology checklist requirements for laboratory accreditation that will become baseline standards for clinical AI deployment.
Reference
| Term | Definition |
|---|---|
| AI pathology | Application of machine learning to analysis of histopathological tissue images |
| WSI | Whole-slide image — a digitised glass slide, typically 1–10 GB per file |
| CPath | Computational pathology — the broader field encompassing AI and quantitative image analysis |
| CNN | Convolutional neural network — the dominant deep learning architecture for image analysis until ~2022 |
| ViT | Vision Transformer — transformer architecture applied to images; dominant in 2023–2026 pathology models |
| MIL | Multiple Instance Learning — technique for slide-level classification without patch-level labels |
| Foundation model | Large model pre-trained on massive unlabelled data, adaptable to downstream tasks |
| FFPE | Formalin-fixed paraffin-embedded — standard tissue preservation method for histopathology |
| H&E | Haematoxylin and eosin — the standard routine tissue stain for histopathological examination |
| IHC | Immunohistochemistry — antibody-based staining to detect specific proteins in tissue sections |
| TIL | Tumour-infiltrating lymphocyte — immune cells within a tumour; assessed as a biomarker |
| Gleason score | Histological grading system for prostate cancer; the primary AI pathology benchmark task |
| ISUP grade | International Society of Urological Pathology grade group — simplified 1–5 prostate grading system |
| MSI | Microsatellite instability — molecular phenotype predictive of immunotherapy response |
| TMB | Tumour mutation burden — total number of somatic mutations; another immunotherapy biomarker |
| HRD | Homologous recombination deficiency — BRCA-related DNA repair defect; predicts PARP inhibitor response |
| SaMD | Software as a Medical Device — regulatory category covering AI diagnostic tools |
| DMS / IMS | Digital / Image Management System — software for storing and organising WSIs |
| LIS | Laboratory Information System — tracks pathology cases, specimens, and reports |
| DINO / DINOv2 | Self-supervised vision ViT pre-training method widely used for pathology foundation models |
| CLIP | Contrastive Language-Image Pre-training — vision-language alignment method (used in PLIP, CONCH) |
| NAS | NASH Activity Score — histological scoring system for non-alcoholic steatohepatitis severity |
| TCGA | The Cancer Genome Atlas — primary public WSI + genomic dataset for model training and benchmarking |
| Stain normalisation | Image processing step to reduce H&E stain variability between labs before AI analysis |
| Federated learning | Distributed model training across multiple sites without sharing patient data |
| CAP | College of American Pathologists — primary US pathology accreditation and standards body |
| De Novo | FDA regulatory pathway for novel medical devices without a predicate; used for Paige Prostate clearance |
Frequently asked questions
Explore Path-iQ
Curated peer-reviewed studies from Nature Medicine, Lancet, JAMA, and Modern Pathology on AI and digital pathology.
Editorial profiles of the world's 100 most influential pathologists across all subspecialties.
Detailed profiles of 100 leading US pathologists at Stanford, Hopkins, Mayo Clinic, and Cleveland Clinic.
Reference directory of the world's leading pathology laboratories and research centres.
Educational and research-focused content only. Not medical advice, diagnosis, or treatment guidance. All regulatory information reflects publicly available data as of July 2026 and may change; verify with manufacturers and regulatory bodies.