Accurate object segmentation using novel active shape and appearance models based on support vector machine learning

This paper presents an accurate object segmentation method using novel active shape and appearance models that evolve according to the output of a support vector machine as well as traditional appearance features at shape landmarks. The method consists of two main processes including the building of the shape and appearance models and support vector machine (SVM) classifier, and the segmentation of test image. In the former process, the shape (or appearance) model is built by extracting the mean shape (or appearance) and a number of modes of variation from training images, and a SVM is trained to classify an image into object pixels or non-object pixels. In the latter process, the predicted object contour (represented with discrete landmarks) starts from the average shape, evolves to a new position as the result of the maximization of the probability of the profile of the landmark being centered at object contour and the minimization of the Mahalanobis distance between a new profile and the profile appearance model, and finally is fitted to the shape model. The method is novel in that the probability of the profile of the landmark being centered at object contour, accurately calculated by SVM classifier, is used in evolving the contour, giving better segmentation result than the original active shape and appearance models where Mahalanobis distance only was used. Experiments of applying the method to segment liver in computed tomography (CT) images were conducted with promising results.