AI better at detecting ovarian cancer

AI-based models are better than human experts at identifying ovarian cancer via ultrasound images. This is the result of a study by researchers at KTH Royal Institute of Technology and KI.
The researchers have developed and evaluated AI models that can distinguish between benign and malignant cell changes in the ovaries. In total, they used over 17,000 ultrasound images from 3,652 patients at 20 hospitals in eight countries to train and test the AI models. The diagnostic capabilities of the AI models were compared against a large group of experts and less experienced ultrasound examiners, all of them medical doctors.
The AI models outperformed both experts and less experienced examiners in identifying ovarian cancer. They had an average accuracy of 86.3 per cent, while the experts scored 82.6 per cent and the less experienced examiners 77.7 per cent.
According to one of the researchers, Emir Konuk , a PhD student in computational science at KTH Royal Institute of Technology, the study shows that AI models can be a valuable aid in diagnosing ovarian cancer. Several other previous studies have also shown this, he says.
"The point of our study is to show that the method works on a broad front with thousands of patients from many different previously unseen hospitals."
The researchers also believe that the AI models can reduce the need for referrals to experts and lead to faster and more cost-effective care for patients with ovarian abnormalities. According to the researchers, further studies are needed to fully understand the potential and limitations of AI models in clinical practice.
In the next step, clinical studies at Södersjukhuset in Stockholm will be conducted to evaluate the safety and usability of AI support in everyday clinical practice.
"With further research as a basis, AI models can play an important role in healthcare and contribute to increased efficiency and higher quality," Emir Konuk says.
He has designed the study's experiments and is the first author together with Filip Christiansen, a doctoral student at KI.
The study, led by researchers at KI, was funded by the Swedish Research Council, the Swedish Cancer Society, Region Stockholm, Radiumhemmets forskningsfond and Wallenberg AI, Autonomous Systems and Software Programme (WASP). The study has been published in the journal Nature Medicine.
Text: Christer Gummeson ( gummeson@kth.se )