New Research Shows AI Can Ask another AI for Second Opinion on Medical Scans

The field of medical artificial intelligence has made remarkable strides thanks to deep learning. However, training these deep-learning models typically requires vast amounts of annotated data. This process of annotating large datasets is not only labor-intensive but also susceptible to human biases, especially for dense prediction tasks like image segmentation. Taking inspiration from semi-supervised algorithms, which utilize both labeled and unlabeled data for training, researchers have created a novel co-training AI algorithm for medical imaging that mimics the process of seeking a second opinion.

The research by scientists at Monash University (Melbourne, VIC, Australia) tackles the challenge of limited availability of human-annotated or labeled medical images by adopting an adversarial, or competitive, learning approach towards unlabeled data. This groundbreaking research is expected to push the boundaries of medical image analysis for radiologists and other healthcare experts. Manually annotating a large number of medical images demands considerable time, effort, and expertise, which often limits the availability of large-scale annotated medical image datasets. The algorithm designed by these researchers enables multiple AI models to harness the unique strengths of both labeled and unlabeled data, learning from each other's predictions to enhance overall accuracy. The next stage of the research will focus on broadening the application to accommodate various types of medical images and developing a dedicated end-to-end product for use in radiology practices.

“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,” said Ph.D. candidate Himashi Peiris of the Faculty of Engineering at Monash University. “This enables AI models to make more informed decisions, validate their initial assessments, and uncover more accurate diagnoses and treatment decisions.”

Related Links:
Monash University