More Accurate Diagnosis of Ovarian Cancer with AI
Region Stockholm
KTH Royal Institute of Technology and Region Stockholm are collaborating to develop an AI support tool that could improve ultrasound diagnostics of ovarian tumors. By using AI as a second reviewer, the researchers aim to reduce unnecessary investigations and help the right patients receive the right care in time.
Every year, around 700 women in Sweden are diagnosed with ovarian cancer. At the same time, ovarian cysts are common – an estimated one in ten women will develop a cyst at some point in their lives. This makes accurate early assessments crucial, so that the right patients can quickly be referred onward, while others can be monitored at the appropriate level of care without causing unnecessary anxiety or burdening specialist healthcare services.
The research project is one of several collaborations in which Region Stockholm and KTH are bringing together clinical practice and technology to solve concrete healthcare challenges. The project has received renewed funding through the joint Health, Medicine and Technology (HMT) initiative, and the results so far suggest that AI could become a valuable tool in the ultrasound diagnosis of ovarian tumors. Together with Associate Professor Pawel Herman and Professor Kevin Smith at KTH, Elisabeth Epstein, professor at Karolinska Institutet and senior consultant at Södersjukhuset, wants to investigate whether AI can strengthen diagnostics by combining ultrasound imaging with advanced blood tests.
“This could provide a more reliable basis for assessment. We also want to explore whether AI can be used to detect ovarian cancer at an earlier stage,” says Elisabeth Epstein.
Large dataset to refine the models
The researchers are currently collecting a large dataset of ultrasound images from the past 15 years, linked to medical records, follow-up data from quality registries, and other examinations. The material is being used to train and fine-tune the AI models so they become more accurate in everyday clinical practice.
According to Pawel Herman, the first milestone is an AI model capable of performing an initial triage – for example, determining whether an image is too poor for analysis and whether a finding should be referred for expert assessment.
“Through deep learning, we want the AI model to recognize cancer while also being able to show how confident the assessment is. AI can support healthcare professionals in different stages of the diagnostic process. At present, the focus is on early detection, where the AI model is designed to carefully evaluate images and avoid missing suspicious cases,” he says.
The importance of trust in decision-making
At the same time, the researchers want to investigate how AI models can be used in a clinical environment and, most importantly, how they can be integrated into the decision-making process. A functioning AI platform is already being tested in clinical studies at Södersjukhuset. Through user studies, the team aims to identify what kind of support doctors need in order to trust AI analyses.
“The machine will not make the decisions – it should support the physician’s decision as a second reviewer. That means it must be clear what the AI’s response is based on, and when further action should be taken,” says Pawel Herman.
Elisabeth Epstein agrees that physicians must feel confident in how to interpret AI responses and know what to do when the AI and the doctor disagree. In addition, the technology must be integrated into healthcare in a way that does not create extra administrative work.
“That is why we are also looking at softer values: how doctors experience working with AI models and how the technology affects their daily work,” she says.
More equal access to ultrasound care
In addition to enabling earlier diagnosis of ovarian cancer, the goal is to introduce AI as a second reviewer more broadly within ultrasound diagnostics, thereby giving all women access to assessments supported by expert-level assistance, according to Elisabeth Epstein.
“By improving the quality of gynecological ultrasound diagnostics, we can achieve more equitable healthcare that ultimately also becomes more cost-effective. We expect the first version of the technology to benefit patients sometime around the turn of 2026/2027. But there is still much to learn about how we can optimize and integrate AI into healthcare. In the long term, we also want to include other gynecological diseases where AI could provide valuable support.”
Both researchers view the collaboration between research and clinical practice as a team effort that has been highly significant to the progress made so far.
“I have collaborated with clinicians in other projects, but the level of commitment and work contributed by the entire team is unique,” says Pawel Herman.
Elisabeth Epstein adds:
“It is exciting to have the opportunity to conduct research with talented people from so many different fields in order to solve an urgent problem – and to be able to take the research all the way to a clinical product.”
Text: Anna-Lena Ahlberg
Photo: Karolinska Institutet, Södersjukhuset and KTH