Why does it matter where a cancer started?
When cancer spreads, cells leave the first organ affected, called the primary tumor, and form new tumors elsewhere. These are metastases. Usually doctors can tell where they came from. But sometimes, even after many tests, the primary tumor cannot be found. This is called a cancer of unknown primary.
That is a real problem, because most modern treatments are chosen according to the organ where the cancer began. A cancer that started in the lung is not treated like one that started in the breast. Genetic tests on the tumor can help find its origin, but they are not always done, and they are rarely available in places with fewer resources.
What did the researchers do?
The team built TOAD, a deep learning AI: a program that teaches itself from thousands of examples. It works on histology slides, which are thin slices of tumor, stained and placed on glass for a pathologist to look at under a microscope. These slides are made for almost every biopsy.
The researchers scanned slides from tumors whose origin was known. They trained the AI to do two things at once: say whether a tumor is primary or a metastasis, and suggest the organ it came from. They then tested it on tumors it had never seen, including an external test set, and on 317 cancers of unknown primary where doctors had reached a working diagnosis.
What were the results?
On the tumors of known origin kept aside for testing, TOAD's first choice was right 83% of the time, and the correct answer was in its top three choices 96% of the time. On the external test set, these figures were 80% and 93%. So the AI did not only work on the kind of images it was trained on.
For the 317 cancers of unknown primary, there is no certain answer to check against. The researchers therefore compared the AI with the doctors' working diagnosis. TOAD's first choice matched it in 61% of cases, and the doctors' diagnosis was among its top three suggestions in 82% of cases.
Key figures
- 83%first choice correct on tumors of known origin (96% within the top three)
- 80%first choice correct on an external test set (93% within the top three)
- 61%agreement with the doctors' diagnosis across 317 cancers of unknown primary (82% within the top three)
Living with a cancer of unknown primary, with Mima
Living with a cancer whose origin is unknown often means waiting a long time for answers. This AI is not used in hospitals yet, but it shows that the biopsy you have already had may hold more clues.
Mima keeps track of your tests and appointments and helps you prepare your questions, such as: "What tests are still possible to find where my cancer started?" You can also tell her how tired you feel from day to day.
What could this mean for patients?
The authors see TOAD as a support tool for difficult cases of metastatic cancer. It could point pathologists towards the right extra tests, or partly stand in for them when those tests are not available. The aim is to reduce the number of cancers whose origin stays unknown, which would help doctors choose a more suitable treatment.
Because the AI only needs the slides already made during a biopsy, it could also help in places where genetic testing is hard to access.
What are the limits of this study?
This is a retrospective study: the AI was tested on slides that already existed, not on patients followed over time. For cancers of unknown primary, the AI was compared with a medical working diagnosis, not with a certain answer. In the remaining cases, its first choice differed from the doctors' view.
The study does not yet show that using TOAD improves treatment or survival. That would need prospective studies, where the tool is used for real in hospitals.
Frequently asked questions
What is a cancer of unknown primary?
It is a cancer found once it has already spread, where tests cannot find the organ it started in. This makes treatment harder to choose, as treatment often depends on that organ.
Does the TOAD AI replace the pathologist?
No. The authors describe it as a support tool for difficult cases. Doctors still make the diagnosis.
What does the AI need to work?
Only scanned images of routine biopsy slides. It does not need a genetic test.
Is this tool used in hospitals?
The study does not show that. It tested the AI on existing slides, and prospective studies are still needed.
The original study
Lu MY, Chen TY, Williamson DFK, et al. “AI-based pathology predicts origins for cancers of unknown primary” Nature, May 5, 2021. DOI : 10.1038/s41586-021-03512-4
Read the full study in NatureThis article summarizes a scientific study and is not medical advice.