Skip to main content

KEY POINTS

  • A new study showcased the application of adversarial learning with AI for medical imaging analysis.
  • The research conducted by Monash University has proposed a novel “dual-view” AI system that could improve the accuracy of image annotations.
  • The algorithm has exhibited enhanced performance in a semi-supervised learning setting, outdoing the latest technologies.

The research published in Nature Machine Intelligence illustrates how AI can be used to improve medical image analysis through adversarial learning. The study, carried out by Monash University, presented a dual-view AI system that enhances the accuracy of image annotations. The AI models can leverage both labeled and unlabeled data to improve overall accuracy. The algorithm outperformed existing methods in semi-supervised learning scenarios, potentially facilitating more accurate diagnoses and treatment decisions.

Designing an Adversarial AI Model for Medical Scans

The research design sought to establish a contest between two components of a dual-view AI system, as explained by Ph.D. candidate Himashi Peiris from the Faculty of Engineering. One segment of the AI system emulates radiologists by labeling medical images, while the second sector evaluates the AI-generated labeled scans by comparing them against the scarce labeled scans offered by radiologists.

Presently, radiologists manually annotate medical scans to identify areas of interest such as tumors. However, this method, being reliant on individual interpretation, is not only time-consuming but also prone to errors and may cause delays for patients awaiting treatment. Moreover, the manual annotation of a large number of images requires significant effort, expertise, and time, hence limiting the availability of large-scale annotated medical image datasets.

The algorithm designed by the Monash researchers allows multiple AI models to utilize the distinct benefits of labeled and unlabeled data, thereby enhancing overall accuracy by learning from each other’s predictions. Peiris shared that “Across three publicly accessible medical datasets, using a 10% labeled data setting, the new algorithm attained an average improvement of 3% when compared to the latest top-performing approach.”

By outpacing preceding technologies, the algorithm offers the potential to facilitate more accurate diagnoses and treatment decisions, even with a limited number of annotations. The next research phase will aim to widen the application scope to include various medical image types and to create an end-to-end product that radiologists can incorporate into their practices.