Researchers at Monash University in Australia have designed a new co-training artificial intelligence (AI) algorithm for medical imaging that can effectively mimic the process of seeking a second opinion. Published recently in Nature Machine Intelligence, the research addressed the limited availability of human annotated, or labeled, medical images by using an adversarial, or competitive, learning approach against unlabeled data.

This research, conducted at Monash University faculties of Engineering and IT, will advance the field of medical image analysis for radiologists and other health experts, says Himashi Peiris, a PhD candidate at the Faculty of Engineering. She further reveals that the research design had set out to create a competition between the two components of a “dual-view” AI system.

“One part of the AI system tries to mimic how radiologists read medical images by labelling them, while the other part of the system judges the quality of the AI-generated labelled scans by benchmarking them against the limited labeled scans provided by radiologists,” says Peiris. 

“Traditionally radiologists and other medical experts annotate, or label, medical scans by hand highlighting specific areas of interest, such as tumours or other lesions. These labels provide guidance or supervision for training AI models,” says Peiris. “This method relies on the subjective interpretation of individuals, is time-consuming and prone to errors and extended waiting periods for patients seeking treatments.”

The availability of large-scale annotated medical image datasets is often limited, as it requires significant effort, time, and expertise to annotate many images manually. The algorithm developed by the Monash researchers allows multiple AI models to leverage the unique advantages of labeled and unlabeled data and learn from each other’s predictions to help improve overall accuracy.

“Across the three publicly accessible medical datasets, utilizing a 10% labeled data setting, we achieved an average improvement of 3% compared to the most recent state-of-the-art approach under identical conditions,” says Peiris.

Plus, she says, “Our algorithm has produced groundbreaking results in semi-supervised learning, surpassing previous state-of-the-art methods. It demonstrates remarkable performance even with limited annotations, unlike algorithms that rely on large volumes of annotated data. This enables AI models to make more informed decisions, validate their initial assessments, and uncover more accurate diagnoses and treatment decisions.”

The next phase of the research will focus on expanding the application to work with different types of medical images and developing a dedicated end-to-end product that radiologists can use in their practices.