AI in Pathology: How Artificial Intelligence Is Changing Cancer Diagnosis

Last updated: October 2, 2026

When doctors suspect cancer, pathology often supplies one of the most important pieces of evidence. A biopsy can show whether abnormal cells are present, what type of tumor may be involved and, often, features that help guide treatment. Traditionally, pathologists examine tissue under a microscope. Increasingly, slides are converted into high-resolution digital images and analyzed with artificial intelligence, a field known as computational pathology or AI-assisted digital pathology.

So what can AI actually do with a cancer biopsy? The answer is more interesting than «it detects cancer.» Modern systems can help identify tumors, classify tissue patterns, quantify biomarkers, measure microscopic features and potentially inform prognosis. But most applications still need careful validation and clinical oversight.

How We Researched This Article

This is a desk-research narrative review. We did not test any software. We:

• Prioritized systematic reviews and peer-reviewed studies from 2024 to 2026, and checked their key figures against the published abstracts.
• Reported numbers as the authors presented them, and noted when a result comes from a selected, retrospective or single-dataset analysis.
• Treated high accuracy in research settings as a starting point, not proof of performance in every hospital.

The Evidence at a Glance

A 2024 systematic review and meta-analysis in npj Digital Medicine screened 2,976 studies, reviewed 100 and pooled 48 in the meta-analysis. It reported high diagnostic accuracy, but 99% of the 100 reviewed studies had at least one area at high or unclear risk of bias or applicability concerns, and the authors called for more rigorous evaluation. That pairing, impressive numbers with shaky study quality, is the central lesson of this field.

What Is Digital Pathology?

A biopsy is processed in the laboratory, embedded, cut into thin sections and stained. A common stain is hematoxylin and eosin (H&E), which makes cell and tissue structures visible. Digital pathology scans the whole slide and creates a whole-slide image (WSI), a very large file that AI can analyze for patterns hard to quantify by hand. The National Cancer Institute describes digital pathology as an end-to-end process in which tissue is digitized and analyzed computationally, with AI models being investigated for cancer diagnosis and other tasks.

How AI Analyzes a Cancer Biopsy

The AI does not necessarily return a single «cancer» or «no cancer» answer. Depending on the system, it may flag suspicious regions, classify tissue, count cells, estimate tumor characteristics or point to areas that deserve a closer look, because pathology asks many different questions:

Different models are built for different tasks.

Where AI Can Help in Cancer Pathology

These applications sit at different stages of development and adoption. A model that scores well on a retrospective dataset is not automatically a clinically validated tool.

1. Detecting Cancerous Tissue

A slide can contain millions of cells. AI can scan the image and highlight regions that match patterns associated with malignancy, which may help when a pathologist faces many slides or when tiny abnormal areas are hard to find. Research has investigated detection in prostate, breast, colorectal and other tumors, but performance varies with cancer type, dataset, slide preparation and intended use. AI is best seen as an extra layer of analysis, not a universal cancer detector.

2. Helping Classify Tumors

Finding cancer is only the beginning; pathologists also need to know what type it is. Models can learn patterns of cells and tissue architecture from large annotated collections. The NCI notes that AI foundation models for digital pathology are being investigated for identifying tumor types and subtypes, and for research into prognosis and treatment response. Instead of only asking «where is the tumor?», researchers now ask what other biological information the tissue image holds.

3. Supporting Tumor Grading

Some cancers are graded by how abnormal they look, which can inform how clinicians understand tumor behavior and plan treatment. AI can quantify certain patterns and may make measurements more reproducible. Prostate cancer has received the most attention, with recent reviews describing AI for diagnosis, grading, prognostic information and molecular subtyping. But grading is not just image recognition: pathologists weigh several features in context, so performance must be tested against clinical standards and across institutions.

4. Measuring Biomarkers

Immunohistochemistry (IHC) uses antibodies to detect specific proteins in tissue and can influence treatment. AI could automate or standardize measuring staining intensity, the share of cells showing a marker and how staining is distributed. Research has examined AI-assisted quantification of markers such as HER2 and PD-L1. The benefit is not only speed: quantitative analysis could reduce some variation between readers, though usefulness depends on validation and clinical context.

5. Reducing Repetitive Work

Pathology involves a lot of counting, measuring and locating. A computer can identify thousands of cells and classify them by set criteria far faster than a person. That does not make the pathologist less important. The work may simply be divided differently: AI handles repetitive image analysis while the pathologist focuses on interpretation, difficult cases and clinical correlation.

6. Additional Prognostic Information

One of the most advanced research areas is estimating prognosis from pathology images, which contain complex patterns that may correlate with tumor behavior. A 2026 systematic review in Cancer Informatics identified 48 studies across 19 cancer types that combined slide images with high-throughput molecular data to predict survival. Every one of them used The Cancer Genome Atlas (TCGA), and only three used it as an independent external dataset. That is a significant limit on how far these models can be assumed to generalize. This is an exciting research direction, but it should not be confused with routine clinical use.

A Real Example: Prostate Cancer

A 2025 study in Communications Medicine tested an AI model on difficult prostate biopsies from three hospitals, all cases where pathologists had already needed IHC staining to reach a diagnosis. The model reached AUROC values of 0.951 to 0.993 for detecting cancer on routine H&E slides. Using sensitivity-prioritized thresholds, it would have reduced the need for IHC staining as shown below, with no false negative predictions.

Why AI Pathology Is Difficult

A digital image sounds ideal for AI, but practice is messier. Hospitals use different scanners, staining protocols, tissue-processing methods and image formats, and small differences in preparation change how tissue looks. Researchers call the resulting problem domain shift: a model trained on one environment can perform differently in another.

Illustration: the same tissue can look different depending on stain, scanner and preparation.

The 2024 review above also noted that important details about case selection, validation and model development were often unclear. That is why accuracy figures should always be read alongside how the model was tested.

AI Does Not Eliminate the Pathologist

An autonomous «AI pathologist» is far simpler than real clinical pathology. A diagnosis may depend on the patient’s history, imaging, laboratory findings and the clinical question behind the biopsy. AI can analyze the image, but the final interpretation has to fit the wider clinical picture. The realistic near-term model is: AI highlights or measures, the pathologist reviews, clinical information is considered, and the diagnosis is finalized. Human oversight is especially important in cancer diagnosis.

What Could Happen Next?

The next phase is likely to go beyond detecting individual cancer cells: systems that analyze several aspects of a slide at once, connect pathology with molecular data and offer prognostic or predictive insights. Foundation models, trained on broad collections of pathology images and adapted to different tasks, are an important research area. A 2026 review described uses from image preprocessing and tumor classification to prognostic stratification and biomarker discovery, while flagging computational demands, noisy annotations, interpretability and differences between datasets. Clinical implementation still requires validation.

The Biggest Questions for Hospitals

Before adopting an AI pathology system, hospitals need to consider more than accuracy. The NCI has highlighted the need to go beyond the algorithm itself and address clinical validation. A practical checklist:

• Has the system been externally validated?
• Was it tested on patients from different hospitals?
• Does performance hold across scanners and staining methods?
• What happens when the AI is wrong?
• Can pathologists understand and review its output?
• How is patient data protected?
• Is the software appropriate for its intended clinical use?
• How will performance be monitored after deployment?

The challenge is no longer just building an accurate model. It is showing that the whole system works safely in a real clinical environment.

Limitations of This Review

• Not a systematic review. We summarized published reviews and studies rather than searching all literature ourselves.
• Retrospective evidence dominates. Much of what is reported comes from retrospective datasets, and results may not carry over to routine care.
• Fast-moving field. Tools and regulatory status change quickly and differ by country.

The Future of AI in Cancer Pathology

AI is turning tissue slides into data that computers can analyze in great detail. It can potentially help pathologists detect tumors, classify cancers, quantify biomarkers, measure tissue characteristics and investigate prognostic patterns. The evidence also shows why caution is needed: high performance in research datasets does not automatically translate to every hospital.

Frequently Asked Questions

Related Reading

Sources

1. McGenity C. et al. Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy. npj Digital Medicine, 2024;7:114. nature.com
2. Kjosavik S.R. et al. Artificial intelligence-assisted prostate cancer diagnosis for reduced use of immunohistochemistry. Communications Medicine, 2025;5:425. nature.com
3. Jennings C. et al. AI in Cancer Prognosis: A Systematic Review of Multimodal Models Combining Pathology Images and High-Throughput Omics. Cancer Informatics, 2026. PubMed
4. National Cancer Institute. AI in digital pathology and biomedical research (overview pages).
5. Zhang et al. Review of AI in digital pathology diagnosis and analysis. Military Medical Research.
6. Kather J.N. et al. Review of AI and digital tools in cancer pathology.