Jiang, Zhifan; Nath, Vishwesh; Roth, Holger R.; Parida, Abhijeet; Foreman, Nicholas; Fisher, Michael J.; Packer, Roger J.; Anwar, Syed Muhammad; Avery, Robert A.; Linguraru, Marius George. “KALM: Knowledge-driven active learning for medical image segmentation using localized similarity.” Proceedings – International Symposium on Biomedical Imaging (2025). https://doi.org/10.1109/ISBI60581.2025.10981164.
When training AI models to identify and outline specific parts of the body in medical images (a process called segmentation), having enough labeled examples is crucial. But labeling medical images takes a lot of time and money because it requires expert input from professionals like radiologists.
One promising solution is active learning—a method that helps AI models learn from fewer labeled images by carefully choosing which unlabeled images would be the most helpful to label next. In this study, we introduce a new active learning approach specifically designed for analyzing 3D medical images.
Our method uses a “selector” that learns from the model’s performance—measured using something called the Dice similarity score—on a set of already labeled images. The selector finds cases where the model struggles, then searches for similar, potentially difficult cases among the unlabeled data. These are then recommended for labeling.
We tested our approach on two different sets of medical images: nearly 500 pediatric brain MRIs (to segment part of the visual system) and 131 contrast-enhanced CT scans (to segment livers and tumors). Our method performed just as well—or better—than other widely-used active learning techniques, while being simpler and more efficient. This shows that our approach can be a powerful tool for making medical AI systems better without needing as much labeled data.
Fig. 1.
T1-weighted MRI sequences with the anterior visual pathway (AVP) ground truth (red) from three sites, demonstrated by the heterogeneity in intensity and a 3D model. CT scan with liver (red) and tumor (green) and a 3D model.
