Segmentation of Femoral Cartilage with a Hybrid Method Combining Voxel Classification and Active Appearance Models

Abstract

In this paper a hybrid segmentation method was proposed for delineation of the femoral cartilage compartment in knee MR images. A formerly developed voxel classification-based segmentation system was combined with an active appearance model (AAM) based segmentation system with an aim to solve the oversegmentation problems of purely classification-based approaches. The voxel classification-based segmentation was achieved through region-growing of sampled voxels depending on one-versus-all classifiers that used approximate nearest neighbour algorithm. Before the appearance model construction, dense set of correspondences were determined on the surfaces of cartilage atlases in 10 training MR images through an iterative shape-context-based non-rigid registration approach. Then, the appearance models for femoral cartilage compartment was constructed either using all of 10 training atlases or grouping these atlases as large and small depending on the physical examination information of the participants. The experimental analyses involved a comparative evaluation of the accuracies of these different AAM-based segmentations both individually and in combination with the voxel classification-based segmentations in 23 testing MR images as well as assessment of the misclassifications of the former segmentation system. As a result, the correspondence finding procedure worked successfully on the training atlases, and the hybrid segmentations with grouped appearance models achieved the closest accuracies to those of the voxel classification-based segmentation. The hybrid segmentations with appearance models that depended on the patient-specific information were evaluated as the most likely method to improve the segmentation accuracies with removal of oversegmented cartilage compartments.