KALM: Knowledge-Driven Active Learning for Medical Image Segmentation Using Localized Similarity

Limited labeled data could compromise the robustness of segmentation models trained on medical images. In medical imaging, manual segmentation is time-consuming and financially costly since expert human labor (radiologist etc.) is required. Active learning (AL) is a promising approach to reduce the amount of labeled data required to train a model by iteratively selecting only the most beneficial instances to be labeled from the pool of unlabeled data. Towards this, we propose a novel AL query paradigm designed for the segmentation of 3D medical images. We use a selector which incorporates the knowledge related to a segmentation model performance measured by the Dice similarity coefficient on a validation dataset. The selector identifies failed validation cases and searches for potentially unsuccessful cases in the unlabeled pool by maximizing a localized image similarity metric. The method is evaluated on two datasets of medical images from multiple sites and modalities: 479 pediatric brain magnetic resonance images for the segmentation of the anterior visual pathway and 131 contrast-enhanced computed tomography scans for liver and tumor segmentation. Our results demonstrate that the proposed AL strategy achieves similar or better segmentation performance than established but computationally more complex uncertainty sampling methods, while showcasing its potential to efficiently select optimal unlabeled data.