Landmark-Free Statistical Shape Modeling Via Neural Flow Deformations

Statistical shape modeling aims at capturing variations of an anatomical structure that occur within a given population. Shape models are employed in many tasks, such as reconstruction and image segmentation, but also generation classification. Existing priors either require dense correspondence between training examples or lack robustness topological guarantees. We present FlowSSM, novel approach learns variability without requiring instances. It relies on hierarchy continuous deformation flows, which parametrized by neural network. Our model outperforms state-of-the-art methods providing expressive robust prior for distal femur liver. show the emerging latent representation is discriminative separating healthy from pathological shapes. Ultimately, we demonstrate its effectiveness two tasks partial data. source code publicly available ( https://github.com/davecasp/flowssm ).