Detecting cancer at an early stage is crucial for effective prevention and treatment. Today, there is another ally that is learning very quickly and becoming increasingly accurate: artificial intelligence. This is the conclusion of a recent study published in Nature Medicine and co-authored by Robert Fruscio, Associate Professor of Gynaecology and Obstetrics at the University of Milano-Bicocca and Director of the Struttura semplice di Ginecologia Preventiva at the IRCCS San Gerardo dei Tintori Foundation: the research, carried out by a team from the Karolinska Institutet in Sweden, involved 20 centres in eight countries and analysed a dataset of more than 17. 000 ultrasound images from more than 3 600 patients, some of whom were referred to the San Gerardo Hospital in Monza. The aim was to train an artificial intelligence programme to distinguish benign from malignant ovarian lesions in these images and to test the potential of these models to support medical diagnosis, reduce the margin of diagnostic error and improve clinical management of patients.
'Ovarian lesions are common and often discovered incidentally, so it is crucial to define their risk of malignancy as accurately as possible in order to guide appropriate treatment,' explains Robert Fruscio. We developed and validated an artificial intelligence system capable of distinguishing between benign and malignant ovarian lesions from an ultrasound image. We then compared the AI's performance with that of experienced ultrasound operators (including myself and other colleagues from around the world) and non-experienced operators. The model proved superior to the experts, albeit by a very small margin, and significantly better than the non-experts. Specifically, the AI-based models achieved an accuracy rate of 86% in detecting ovarian cancer, compared to 82% for human experts and 77% for those with less experience. The results were consistent regardless of the age of the patients, the ultrasound equipment used and the clinical context.
The importance of this study lies in the general context that experienced operators are in short supply in many parts of the world and are not available in all hospitals. The shortage of experienced sonographers leads to unnecessary surgery and delayed diagnosis of cancer. Artificial intelligence models could therefore help less experienced operators to select patients for referral to second-tier centres and avoid unnecessary surgery in patients with low-risk lesions," continues Robert Fruscio. Overall, this is a classic case where AI will not replace humans, but could improve the efficiency of the whole system and the management of patients.
The study also found that in a triage simulation, AI-assisted diagnosis would reduce referrals to specialists by 63%, significantly outperforming current in-office diagnosis. While the study emphasises that further prospective and randomised trials are needed to validate the clinical benefits and cost-effectiveness of AI-assisted diagnosis, it also highlights the need for further research.