Knowledge-based Active Appearance Model applied in medical image localization

Active Appearance Model (AAM) is a statistical model of object localization. Based on AAM, researchers have attained good performance on local object localization; however, they haven't attached much importance to the correlations among different objects in an image. To address this issue, our paper presents a knowledge-based AAM. With X-ray medical images of human spines, we first apply AAM to middle joint of the spine. Then we rely on spatial correlations between this middle joint and the neighboring ones to adjust mean model for localization of neighboring joints. Concerning sensitivity of AAM to different starting positions, we propose a histogram-based approach for locating rough starting position. By comparing our experimental results with different models, we conclude that the knowledge-based AAM is efficient, accurate and robust.