Active Learning with nnUNet for Coronary Artery Lumen Segmentation Using a Centerline Prior

Annotating medical images for segmentation is both costly and time-consuming, making it crucial to identify the most informative images for annotation. Active learning aims to address this challenge by selecting samples that maximize model performance while minimizing labeling effort. This paper presents an active learning framework that incorporates an anatomical prior for coronary artery segmentation, using nnUNet as the segmentation model. We introduce two novel centerline-based sampling strategies, Lowest Weighted Overlap (LWOV) and Highest Weighted Overlap (HWOV), designed to enhance structural consistency in model predictions. The method is evaluated on Left Anterior Descending (LAD) artery segmentation from Computed Tomography (CT) images. Our results show that although all the active learning strategies evaluated performed well with marginal differences, random sampling achieved the highest performance, highlighting the challenges of designing optimal selection strategies. Furthermore, we demonstrate that with only 16.6% of the available data, we achieve segmentation accuracy comparable to training on the full dataset.