Shape Models from Examples using Automatic Shape Clustering andProcrustes

1. Original contribution of this work: Development of a shape classiication method. Development of a novel approach to aligning arbitrary shapes in a common coordinate-parameterization frame. Accuracy improvements of current 2-D exible shape registration methods. Application of the developed methods to brain model design and MR brain image segmentation. 2. Importance of the contribution: Shape learning is one of the open problems in computer vision. Deformable template segmentation using a learned shape model is of primary importance in medical image analysis. Our novel approach facilitates automated learning of 2D shapes and provides a convenient way to derive and train shape models. and others in the area of shape representation and registration is also highly relevant. Our shape learning method substantially diiers from those of Bookstein and Wolfson Image Analysis Group. The main diierence is that not all the shapes provided in the training set are used for computing the average. The training set is rst automatically clustered and those shapes considered to be outliers are discarded. The second diierence is in the way how well-registered sets of points of the same cardinality are extracted from each shape contour. 4. Interest to others: There is a pressing need to develop automated approaches to many medical image analysis tasks. In most applications, the shape models actually used are derived either by aligning the training data manually and semi-automatically or by bypassing the alignment step. We think that a good model should separate shape information from pose and parameterization information. On the other hand, human intervention in the training process is time consuming, tedious and at many times irreproducible. Our method could be used as an initial alignment step by all shape learning algorithms independently of the shape representation used. Though we report results only on brain structures, our method is generally applicable to many tasks involving deformable shape analysis. The method is powerful, robust, widely applicable, and easy to implement. Abstract A new fully automated shape learning method is presented. It is based on clustering a shape training set in the original shape space and performing a Procrustes analysis on each cluster to obtain a cluster prototype and information about shape variation. Our shape learning method substantially diiers from the previously reported ones. The main diierence is that not all the shapes provided in the training set are used for computing the average. The training set is rst automatically clustered and those …