In this paper, we present a three-dimensional interactive segmentation method. Unlike most previous interactive
methods which largely depend on user interaction, we exploit a prior knowledge of training data to reduce the
user effort. Based on the prior knowledge, most distinguishable parts of an object are automatically segmented
and labels of some uncertain parts are queried to an user. To systematically model the problem, we combine the
hierarchical Markov random field (HMRF) framework and the active learning scheme. The HMRF framework,
proposed for the automatic manner, simultaneously reflects characteristics of local variations and their global
smoothness, while the active learning scheme improves the efficiency of interactive system. We incorporate the
active learning strategy into the editing step of the HMRF structure in order to find and modify the uncertain
parts after the automatic segmentation. Specifically, the uncertainties of local regions are firstly computed by the
label difference between segmentation candidates. Then, the graph models of the uncertain regions are updated
by the user interaction. Since the HMRF structure constrains the smoothness of local regions and the global
optimality, the segmentation is updated as a whole even though the small numbers of local parts are edited.
The proposed method is applied to the segmentation of femur and tibia in knee MR images for evaluation. The
evaluation demonstrates that the proposed method improves the segmentation efficiency more than the graph
cut based method or manual editing.