Why are ovarian masses hard to assess?
Ovarian masses, such as cysts, are common. They are often found by chance, during a scan done for another reason. Most are not cancer, but some are.
Doctors use ultrasound, a painless scan that uses sound waves, to decide. The problem is that reading these images well takes a lot of experience, and there are not enough expert examiners. The authors say this shortage raises two worries: unnecessary operations for harmless masses, and cancers diagnosed too late.
What did the researchers do?
An international team, led from Sweden, gathered 17,119 ultrasound images from 3,652 patients at 20 centers in eight countries. The images came from 21 different scanners. For each patient, the final diagnosis was known, usually from testing the mass after surgery.
They trained AI models called transformers, a type of neural network widely used today. To check that the AI works beyond the places where it learned, they tested it center by center: each time it learned from 19 centers and was tested on the 20th, which it had never seen. Its answers were compared with those of human examiners, both experts and non-experts.
The AI outperformed expert examiners
On unseen centers, the AI reached an F1 score of 83.5%. This score sums up, in one number, how well a test finds cancers without raising too many false alarms. Expert examiners scored 79.5% and non-experts 74.1%. The AI did better on every measure, including sensitivity (cancers found) and specificity (harmless masses correctly recognized).
Its results held up across centers, scanners, types of mass and patient ages. In a simulation where the AI helped non-expert examiners decide, referrals to experts fell by 63%, while diagnosis was still more accurate than current practice.
Key figures
- 17,119ultrasound images from 3,652 patients at 20 centers in eight countries
- 83.5%F1 score for the AI on unseen centers, higher than both expert and non-expert examiners
- 63%fewer referrals to experts in a triage simulation
Keeping track of an ovarian mass with Mima
This AI is not used routinely yet: it read past ultrasound scans. But it shows how much an expert opinion matters, and how hard the wait between finding a mass and getting a diagnosis can be.
With Mima, you can keep track of your scans, their reports and your appointments. She also helps you prepare your questions, for example: "Do I need a specialist opinion for this mass, and how soon?"
What this could change for patients
Today, a woman with an ovarian mass may wait a long time for an expert opinion, or have an operation that turns out to be unnecessary. An AI at this level could help the first doctor sort cases better: reassure sooner when a mass looks harmless, and refer quickly to a specialist when it looks suspicious.
It could also ease the load on experts, who are in short supply, so they can focus on the hardest cases. The authors believe their models are ready for clinical implementation studies.
What are the limits of this study?
This is a retrospective study: the AI read scans that had already been done, and it was not used live with patients. The authors write that prospective and randomized studies are still needed to prove a real benefit.
The main comparison only included women who had surgery, with a diagnosis confirmed under the microscope. Results could differ for women whose masses are simply monitored with follow-up scans. And the triage simulation is still a simulation, not a working hospital service.
Frequently asked questions
Can AI tell whether an ovarian cyst is cancer?
In this study, an AI told benign and malignant masses apart on ultrasound better than experts. But it read past scans: it is not yet used routinely to make a diagnosis.
Are most ovarian cysts cancer?
No. Ovarian masses are common and most are benign. The challenge is to spot the ones that are not, without operating on the others for nothing.
Will AI replace the person doing the scan?
No. The authors see it as a diagnostic aid, for example to help a non-expert doctor decide whether to ask for an expert opinion.
When will this AI be available?
The authors believe it is ready for clinical studies. Prospective trials must first show a real benefit for patients.
The original study
Christiansen F, Konuk E, Ganeshan AR, et al. “International multicenter validation of AI-driven ultrasound detection of ovarian cancer” Nature Medicine, January 2, 2025. DOI : 10.1038/s41591-024-03329-4
Read the full study in Nature MedicineThis article summarizes a scientific study and is not medical advice.