Weakly supervised 3D reconstruction of the knee joint from MR images using a volumetric Active Appearance Model

Active Appearance Models (AAM), have been widely used for the segmentation of anatomical structures in 3D medical images. Building the AAM usually requires the manual segmentation of a training set. In this paper, we propose to reduce this manual segmentation by building a volumetric AAM (vAAM) from MR images. These images are converted to a tetrahedral mesh representation rather than a voxel representation. Tetrahedral meshes are generated so that they represent the underlying image structures. The vAAM is iteratively built using the minimum description length principle (MDL). The generated model provides an anatomical correspondence between the tetrahedral meshes that are generated from MR images within the training set. Thus any manual segmentation performed on a single MR image can be mapped to other MR images. The segmentation of a query image is performed automatically by adapting the vAAM to this image. After the segmentation step, the 3D reconstruction of the knee surface is simply performed by extracting faces that are shared by adjacent regions.