Shape modeling with few training data

Generative statistical shape models have a long and successful history in medical image analysis as priors for image segmentation or registration tasks. Nowadays they are also successfully utilized within deep learning approaches where they either serve as strong prior knowledge or are being used during model training to generate additional artificial but realistic training data. A key problem of traditional statistical shape models is their inability to generalize well to unseen new data when they were trained with few samples. Resulting models will mostly focus on global shape relations and not be able to model the true variability to be expected in a large population. To alleviate this problem, we develop shape modeling approaches that use data-driven and physics-inspired mechanisms during training to create more flexible models even in small training data regimes. Key contributions are the introduction of the principle of locality, which breaks up long-range relations in the training data, and kernel-based techniques to allow the integration of this principle into medical image analysis frameworks that use large 3D data.