Active learning reduces annotation effort in medical image segmentation, yet strategy comparisons typically focus on global metrics such as the Dice coefficient while overlooking which images are acquired and how acquisition patterns influence performance. We present a strategy-centered analysis of active learning for lung segmentation on the JSRT dataset (247 chest Xrays), treating acquisition decisions as first-class objects of study. Three strategies are compared: Random, SDS (Segmentation Dispersion Score), and CAS (Composite Acquisition Score), using U-Net with MC-dropout. Going beyond Dice, we additionally evaluate boundary precision via HD95 and report strategydependent acquisition patterns, outlining a plan to establish whether acquisition diversity correlates with boundary quality and reliability.