Detecting anatomical structures, such as the carina, the pulmonary trunk and the aortic arch, is an important step in designing a CAD system of detection Pulmonary Embolism. The presented CAD system gets rid of the high-level prior defined knowledge to become a system which can easily extend to detect other anatomic structures. The system is based on a machine learning algorithm --- AdaBoost and a general feature --- Haar. This study emphasizes on off-line and on-line AdaBoost learning. And in on-line AdaBoost, the thesis further deals with extremely imbalanced condition. The thesis first reviews several knowledge-based detection methods, which are relied on human being's understanding of the relationship between anatomic structures. Then the thesis introduces a classic off-line AdaBoost learning. The thesis applies different cascading scheme, namely multi-exit cascading scheme. The comparison between the two methods will be provided and discussed. Both of the off-line AdaBoost methods have problems in memory usage and time consuming. Off-line AdaBoost methods need to store all the training samples and the dataset need to be set before training. The dataset cannot be enlarged dynamically. Different training dataset requires retraining the whole process. The retraining is very time consuming and even not realistic. To deal with the shortcomings of off-line learning, the study exploited on-line AdaBoost learning approach. The thesis proposed a novel pool based on-line method with Kalman filters and histogram to better represent the distribution of the samples' weight. Analysis of the performance, the stability and the computational complexity will be provided in the thesis. Furthermore, the original on-line AdaBoost performs badly in imbalanced conditions, which occur frequently in medical image processing. In image dataset, positive samples are limited and negative samples are countless. A novel Self-Adaptive Asymmetric On-line Boosting method is presented. The method utilized a new asymmetric loss criterion with self-adaptability according to the ratio of exposed positive and negative samples and it has an advanced rule to update sample's importance weight taking account of both classification result and sample's label. Compared to traditional on-line AdaBoost Learning method, the new method can achieve far more accuracy in imbalanced conditions.