Researchers at Monash University have developed a new co-training AI algorithm for medical imaging that can imitate the process of seeking a second opinion.
Recently published in Nature Machine Intelligence, the study addressed the limited availability of annotated medical images by utilizing an adversarial learning approach against unlabelled data.
This research, conducted by the faculties of Engineering and IT at Monash University, will advance medical image analysis for radiologists and other health experts.
Himashi Peiris, a PhD candidate from the Faculty of Engineering, explained that the research aimed to create a competition between the two components of a “dual-view” AI system.
According to Ms Peiris, “One part of the AI system attempts to mimic how radiologists interpret medical images by labeling them, while the other part of the system assesses the quality of the AI-generated labeled scans by comparing them to the limited labeled scans provided by radiologists.”
She further added, “Typically, radiologists and other medical experts manually annotate medical scans by highlighting areas of interest, such as tumors or lesions. These labels guide the training of AI models. However, this method is subjective, time-consuming, prone to errors, and causes extended waiting times for patients seeking treatment.”
Annotated medical image datasets are often scarce due to the significant effort, time, and expertise required for manual annotation.
The algorithm developed by the Monash researchers enables multiple AI models to leverage the advantages of labeled and unlabeled data and learn from each other’s predictions to improve overall accuracy.
“In three publicly accessible medical datasets, using 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,” Ms Peiris said.
“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 allows AI models to make more informed decisions, validate initial assessments, and uncover more accurate diagnoses and treatment decisions.
The next phase of the research will focus on expanding the application to different types of medical images and developing a dedicated end-to-end product that radiologists can use in their practices.
The study, led by Associate Professor Mehrtash Harandi and conducted by Himashi Peiris, Associate Professor Zhaolin Chen, Dr Munawar Hayat, and Professor Gary Egan, was published in Nature Machine Intelligence and involved Monash University’s Faculty of Engineering, Monash Biomedical Imaging, and the Faculty of Information Technology.