Why do pathology AIs struggle to work everywhere?
To diagnose a cancer, a specialist doctor (the pathologist) looks under a microscope at a thin, stained slice of tissue, the biopsy slide. Today these slides are scanned into very detailed images, so an AI can analyse them.
The problem is that most AIs are built for a single task, such as spotting one type of cancer. They often work less well when images come from another hospital, another scanner or another group of patients.
What did the researchers do?
The team built CHIEF, a "foundation model": a general-purpose AI that first learns to recognise patterns across a huge number of images before being adapted to specific tasks. It was trained on 60,530 slides covering 19 parts of the body, which is 44 terabytes of images.
It learned in two ways: first from small tiles of an image, without being told what it was looking at, then from whole slides with broad hints (called "weakly supervised" learning).
What were the results?
CHIEF was used to detect cancer cells, identify a tumour's origin, describe a cancer's molecular profile and estimate outcomes. To check that it does not only work on images it has already seen, the researchers tested it on 19,491 slides from 32 independent slide sets, gathered from 24 hospitals and cohorts around the world.
Overall, CHIEF beat the best deep learning methods it was compared with, by up to 36.1%. The authors stress that it stays effective on samples from different populations and prepared in different ways.
Key figures
- 60,530digitised biopsy slides used to train CHIEF, covering 19 parts of the body
- 19,491test slides from 32 independent sets and 24 hospitals and cohorts
- up to 36.1%better than the best AI methods it was compared with
What does this change for patients?
In time, a tool like this could help pathologists save time and get more information from the same slide, for example about the origin of a hard-to-classify tumour. It could also offer a second opinion where expertise is scarce.
For now, nothing in the study shows a benefit for patients followed in real life. CHIEF is a research result, not an approved diagnostic tool.
What are the limits?
Performance was measured on existing datasets, not in a trial where the AI changed care decisions. The authors describe a foundation for more efficient slide evaluation, not a replacement for the pathologist.
The tool still has to be tested in routine practice, checked to make sure it works fairly for all patient groups, and approved before any medical use.
Biopsy: understanding your report with Mima
CHIEF is still a research tool and is not used to read your biopsy in routine care. One day it could help pathologists save time and get more information.
With Mima, you can keep your report and appointments in one place, and have the technical words explained. A concrete question: "Does my tumour need further tests?"
Frequently asked questions
What is CHIEF?
CHIEF is a general-purpose AI that analyses digitised images of cancer biopsies. It was described in Nature in 2024.
Can an AI replace the pathologist?
The study does not show that. CHIEF is presented as a foundation for more efficient slide evaluation, with the pathologist still making decisions.
What is a foundation model?
It is an AI first trained on a very large amount of data to learn general patterns, then adapted to several specific tasks.
Is this tool already used in hospitals?
The paper reports a research result. It does not describe an approval or routine use.
The original study
Wang X, Zhao J, Marostica E, Yu KH, et al. “A pathology foundation model for cancer diagnosis and prognosis prediction.” Nature, 4 September 2024. DOI : 10.1038/s41586-024-07894-z
Read the full study in NatureThis article summarises a scientific study and is not medical advice.