A new international study led by researchers at Karolinska Institutet in Sweden shows that AI-based models can outperform human experts at identifying ovarian cancer in ultrasound images. The study is published in Nature Medicine.
“Ovarian tumors are common and are often detected by chance,” said Professor Elisabeth Epstein at the Department of Clinical Science and Education, Södersjukhuset (Stockholm South General Hospital), at Karolinska Institutet and Senior Consultant at the hospital’s Department of Obstetrics and Gynecology. “There is a serious shortage of ultrasound experts in many parts of the world, which has raised concerns of unnecessary interventions and delayed cancer diagnoses. We therefore wanted to find out if AI can complement human experts.”
AI Outperforms Experts
The researchers have developed and validated neural network models able to differentiate between benign and malignant ovarian lesions, having trained and tested the AI on over 17,000 ultrasound images from 3,652 patients across 20 hospitals in 8 countries. They then compared the models’ diagnostic capacity with a large group of experts and less experienced ultrasound examiners.
The results showed that the AI models outperformed both expert and nonexpert examiners at identifying ovarian cancer, achieving an accuracy rate of 86.3%, compared to 82.6% and 77.7% for the expert and nonexpert examiners respectively.
“This suggests that neural network models can offer valuable support in the diagnosis of ovarian cancer, especially in difficult-to-diagnose cases and in settings where there’s a shortage of ultrasound experts,” said Professor Epstein.
Reducing the Need for Expert Referrals
The AI models can also reduce the need for expert referrals. In a simulated triage situation, the AI support cut the number of referrals by 63% and the misdiagnosis rate by 18%. This can lead to faster and more cost-effective care for patients with ovarian lesions.
Despite the promising results, the researchers stress that further studies are needed before the full potential of the neural network models and their clinical limitations are fully understood.
“With continued research and development, AI-based tools can be an integral part of tomorrow’s health care, relieving experts and optimizing hospital resources, but we need to make sure that they can be adapted to different clinical environments and patient groups,” said Filip Christiansen, a doctoral student in Professor Epstein’s research group at Karolinska Institutet and joint first author with Emir Konuk at the KTH Royal Institute of Technology.
Evaluating the Safety of the AI Support
The researchers are now conducting prospective clinical studies at Södersjukhuset to evaluate the everyday clinical safety and usefulness of the AI tool. Future research will also include a randomised multicentre study to examine its effect on patient management and health-care costs.
Disclosures: The study was conducted in close collaboration with researchers at the KTH Royal Institute of Technology and was financed by grants from the Swedish Research Council, the Swedish Cancer Society, the Stockholm Regional Council, the Cancer Research Funds of Radiumhemmet and the Wallenberg AI, Autonomous Systems and Software Program (WASP). For full disclosures of all study authors, visit nature.com.
