This paper presents a new active shape models(ASMs) based method to segment the prostate from CT images for the radiotherapy. The key point of ASMs is the construction of both shape model and appearance model. We utilize the scale invariant feature transform(SIFT) local descriptor， which is more distinctive than general intensity and gradient features on the edges of the prostate boundary in the CT images， to characterize the image features and build the appearance model. To accurately capture prostate shape variation， an online training mechanism is proposed to build the shape model. When the samples of current patient are limited， the population statistics is used to build the shape model. As the increase of the samples of current patient， the patient-specific statistics plays an important role for constructing the shape model gradually. We test our method on a data set including 264 images of 24 patients， the average Dice similarity coefficient (DSC) is 90.5% and the mean average surface distance(ASD) is 1.90mm. The results show that the proposed method is robust and accurate.