Why are rare cancers so hard for AI?
To diagnose cancer, a specialist doctor called a pathologist looks at thin, stained slices of tissue under a microscope. These are called slides. More and more hospitals now scan their slides, so an AI can analyze them.
Until now, most pathology AI tools have been trained for a single organ, using thousands of examples labeled by hand. That works for common cancers, such as breast or prostate cancer. But for a rare cancer, which the National Cancer Institute (NCI) defines as one affecting fewer than 15 in 100,000 people a year in the US, there are not enough examples to build a dedicated tool.
What did the researchers do?
The team, mostly from the company Paige, working with Microsoft Research and Memorial Sloan Kettering, built Virchow, named after the doctor seen as the father of modern pathology. It is a foundation model: it first teaches itself, without labels, to describe what it sees on 1,488,550 slides from 119,629 patients. Those descriptions then become the base for more focused tools, which need far fewer labeled examples.
On top of this base, the researchers built a single cancer detector for all tissues. They tested it on 9 common cancers (breast, prostate, lung, colon, skin, bladder, uterus, pancreas, head and neck) and 7 rare ones (liver, stomach, brain, ovary, cervix, testis, bone). Half of the test samples came from other institutions, sent in for a second opinion.
What were the results?
Results are reported as area under the curve, a score from 0.5 (pure chance) to 1 (a perfect test). The detector scored 0.950 across all cancers and 0.937 on rare cancers, better than the other foundation models it was compared with. When set to catch 95% of cancers, it correctly cleared 72.5% of samples without cancer.
The researchers also compared it with three commercial Paige tools, each trained for one organ. On prostate and breast, it fell slightly short (0.980 vs 0.995 for prostate). But it did better on some rare forms of breast cancer, and it even detected lymphomas, cancers of the lymph system that none of these models had been trained to find.
Key figures
- 1.5 milliondigital biopsy slides used to train Virchow
- 0.950area under the curve for detecting cancer, all types combined
- 0.937area under the curve for the 7 rare cancers
Living with a rare cancer: staying organized with Mima
Virchow is still a research tool. But with a rare cancer, getting a diagnosis can take time and may involve a second opinion at a specialist center, like the cases in this study.
Mima helps you keep track of your biopsies, reports, and appointments, and prepare your questions, such as: "Can my slides be reviewed by a center that specializes in my cancer?" You can also connect with other patients going through the same thing.
What could this mean for patients?
Most biopsies turn out to be benign. According to the authors, a detector like this could help pathologists look at suspicious slides first and shorten the wait for results. And because it needs fewer examples to train, reliable tools could finally be built for rare cancers, which AI often leaves behind.
Either way, the AI would be an assistant: the pathologist still makes the diagnosis.
What are the limits of this study?
The training slides came from one center and a small number of scanner types. Results varied by cancer: bone cancer (0.841) and cervical cancer stayed below 0.9. More than half of the false alarms were precancerous changes rather than healthy tissue.
The authors say detailed validation, cancer type by cancer type, is still needed before clinical use. Most of them work for Paige, which makes commercial pathology products. Finally, the study used existing slides, so it does not yet show any benefit for patients.
Frequently asked questions
What is a foundation model in pathology?
It is a general-purpose AI trained on a huge number of unlabeled slides. It then serves as the base for more specific tools, such as a cancer detector, that need only a few examples.
What counts as a rare cancer?
The National Cancer Institute defines a rare cancer as one that affects fewer than 15 in 100,000 people a year in the US. In this study, they included cancers of the liver, brain, ovary, and bone.
Will AI replace pathologists?
No. Virchow is a research tool. The authors see it as a way to help sort and prioritize slides, with the pathologist still making the diagnosis.
Is Virchow used in hospitals today?
The study tested it on existing slides and does not describe routine use. Detailed validation for each cancer type is still needed.
The original study
Vorontsov E, Bozkurt A, Casson A, et al. “A foundation model for clinical-grade computational pathology and rare cancers detection” Nature Medicine, July 22, 2024. DOI : 10.1038/s41591-024-03141-0
Read the full study in Nature MedicineThis article summarizes a scientific study and is not medical advice.