Less Is More: A Comparison of Active Learning Strategies for 3D Medical Image Segmentation