AI model predicts 32 cancer tumor types from images already used in hospitals

A routine cancer slide may hold more information than doctors can see through a microscope. A new study suggests artificial intelligence can read those familiar tissue images and predict cancer type, key genetic changes and some survival-related outcomes across 32 solid tumors.

The research points to a growing link between pathology and molecular cancer care. The AI model studied ordinary hematoxylin and eosin-stained slides, known as H&E slides. These slides already guide cancer diagnosis in hospitals around the world.

Researchers trained the model using more than 11,000 primary tumor cases from the Pan-Cancer Atlas. The data included whole slide images, mutation records, RNA-sequencing results and clinical outcomes.

Histopathology remains the gold standard for diagnosing most solid tumors. Pathologists examine thin tissue sections under a microscope and look for tumor shape, structure and abnormal cell patterns.

Yet modern cancer care often needs more than visual diagnosis. Doctors may also need molecular tests to identify genetic changes that affect tumor growth, treatment response and long-term risk.

One major gene is TP53, one of the most frequently altered tumor suppressor genes in human cancer. When TP53 works properly, it helps regulate cell growth, repair DNA damage and remove badly damaged cells. When it mutates, cancer can become more aggressive.

“Standard molecular profiling for TP53 mutations is often costly and inaccessible in underprivileged or remote clinical settings,” said co-lead investigator Alex W. Hewitt, PhD, of the Menzies Institute for Medical Research and School of Medicine, University of Tasmania.

Most AI tools in pathology focus on one task at a time. One model may classify tumor type. Another may predict a mutation. A third may estimate risk.

This study took a broader approach. The researchers built a single Vision Transformer-based model that could generate seven outputs from one whole slide image.

“We wanted to develop a more practical tool for pathologists. Currently, most deep learning-based models are used for single-model concepts; one model for one task. We developed a single model that can generate seven outputs simultaneously from the whole histopathology image, including TP53 mutation status, TP53 RNA expression, tumor type, and survival-related outcomes at the slide level,” Hewitt said.

That multitask design matters because cancer diagnosis already involves many layers of information. A tool that gathers several clues from one slide could fit more naturally into clinical workflows.

Whole slide images are enormous and complex. A single slide may contain tumor, normal tissue, empty background, blur, stains and artifacts.

Experts cannot easily label every important region by hand. That process takes time, costs money and can vary between pathologists.

To handle this challenge, the researchers used weakly supervised learning. This means the model learned from slide-level labels rather than detailed pixel-by-pixel markings.

“In this study, molecular labels such as TP53 mutation status were available at the patch level, but whole slide images containing TP53-associated morphological information were not manually labeled,” Hewitt said.

He said weak supervision allowed the model to identify relevant tissue patterns without exhaustive manual annotation.

The model broke each slide into smaller image patches. It then searched those patches for patterns linked to cancer type, TP53 mutation status and TP53 RNA expression.

The AI used a Vision Transformer, a model designed to find relationships across complex images. It combined many patch-level signals into a slide-level prediction.

The strongest headline result involved TP53 mutation detection. In an independent validation set of 1,729 slides, the model reached an AUROC of 0.766 across 32 solid tumor types.

AUROC measures how well a model separates positive and negative cases. A score of 0.5 means chance-level performance. A score closer to 1.0 means stronger prediction.

The model also inferred TP53 RNA expression levels and tumor taxonomy directly from whole slide images.

TP53 is often called a guardian gene because it helps protect cells from dangerous damage. When this gene changes, cells may lose important growth controls.

These mutations can affect tumor behavior, treatment resistance and disease progression. That makes TP53 status valuable for understanding cancer biology.

In many hospitals, molecular profiling helps doctors see these changes clearly. But access to testing differs widely. Some clinics may lack advanced genomic tools, especially in remote or under-resourced areas.

AI will not replace molecular testing. But it may help decide which patients most need confirmatory testing.

“This approach could help identify patients who may benefit from confirmatory molecular testing, support triage in settings with limited genomic testing, and provide additional decision support to clinicians,” said co-lead investigator Abadh K. Chaurasia, PhD, of the Menzies Institute for Medical Research, University of Tasmania, and Pandani Solutions Pty Ltd.

The researchers stressed that the tool should support existing diagnostic pathways, not replace them. Molecular tests remain the standard for confirming genetic alterations.

“Importantly, this method should be viewed as complementary to molecular testing, not a replacement,” Chaurasia said. “Its potential impact is strongest as a screening, prioritization, or decision-support tool within broader diagnostic pathways.”

That caution matters. AI predictions can help flag likely cases, but cancer care needs high confidence. Treatment decisions should rely on validated clinical evidence.

The model’s performance was strongest for cancer classification and TP53-related prediction. Survival-related prediction remained more limited, which reflects the complexity of patient outcomes.

Survival depends on tumor stage, treatment, general health, follow-up and many other factors. A slide can reveal much, but not everything.

The study shows how computational pathology may help connect standard images with deeper cancer biology. A slide prepared for diagnosis could also offer clues about molecular changes.

This could help doctors make better use of tissue they already collect. It could also reduce delays in deciding which cases need more advanced testing.

The approach may be especially useful in health systems with limited genomic testing capacity. In those settings, AI-assisted triage could help prioritize scarce resources.

He concludes, “Accurate molecular profiling from routine histopathology slides, already widely used in cancer care, could transform clinical oncology. This new AI-based model integrates diagnostic, molecular, and prognostic tasks, and could help clinicians obtain more information from existing pathology workflows, ultimately supporting more accessible precision cancer care and early intervention.”

This research could help make precision cancer care more accessible. If validated in future studies, AI tools may help doctors extract more information from routine slides already used in diagnosis. That could reduce delays and help guide which patients need molecular testing first.

The findings may also support hospitals with limited access to genomic testing. In under-resourced or remote settings, an AI-based screening tool could help prioritize confirmatory tests for patients most likely to benefit. This would not replace sequencing, but it could make testing pathways more efficient.

For researchers, the study advances computational pathology across many tumor types. Instead of building separate models for every cancer and every task, scientists may develop broader systems that connect tissue structure with molecular and clinical data.

For patients, the long-term promise is more complete information from the first diagnostic slide. Cancer care could become faster, more targeted and more equitable if tools like this prove reliable in real-world clinics.

These resources explore how AI can extract diagnostic, genetic and prognostic information from routine pathology slides, including efforts to move these systems toward real-world cancer care.

End-to-end multimodal pathology foundation model with clinical dialogue
PRISM2 was trained on 2.3 million whole-slide images and pathology-report data, then tested across cancer detection, biomarker and survival tasks. It shows how large pathology models can combine tissue morphology with clinical information while supporting several types of prediction from the same system. (Nature Medicine, 2026)

A multimodal whole-slide foundation model for pathology
This study introduces TITAN, a foundation model built to analyze entire pathology slides rather than isolated image patches. The model was evaluated across diagnostic classification, rare-cancer retrieval and other pathology tasks, illustrating the broader move toward general-purpose AI systems that can extract multiple kinds of information from tissue images. (Nature Medicine, 2025)

Real-world deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection
Researchers used routine H&E slides to predict EGFR mutations in lung adenocarcinoma and tested the model prospectively in a clinical setting. The AI-assisted workflow reduced the need for rapid molecular tests by as much as 43% while maintaining the existing clinical performance standard. (Nature Medicine, 2025)

Deep learning using histological images for gene mutation prediction in lung cancer: a multicentre retrospective study
The DeepGEM model analyzed routine lung cancer pathology images to predict several driver mutations across data from 16 hospitals and external datasets. The work provides multicenter evidence that visible tissue patterns can carry information linked to mutations such as EGFR and KRAS. (The Lancet Oncology, 2025)

A whole-slide foundation model for digital pathology from real-world data
Prov-GigaPath was trained on 1.3 billion image tiles from more than 171,000 pathology slides and performed strongly across cancer subtyping and molecular prediction tasks. Its results included improved prediction of multiple cancer-associated mutations, demonstrating how large-scale whole-slide models can connect routine tissue morphology with genomic information. (Nature, 2024)

Research findings are available online in the journal American Journal Of Pathology.

The original story "AI model predicts 32 cancer tumor types from images already used in hospitals" is published in The Brighter Side of News.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.