Learning skull shape variations through CT scan data

초록

We address the problem of inferring a plausible skull geometry from the visible facial surface of a human head. This capability is valuable for medical applications, as it enables skull visualization without exposing patients to ionizing radiation from CT imaging. Unlike CT scans, parametric shape models provide flexible, low-dimensional control for intuitive shape editing. However, their usage typically requires expertise in 3D graphics. To overcome this limitation, we propose a method for predicting a plausible skull shape directly from a single RGB image. Our approach employs an unlabeled registration strategy, making it scalable to clinical data. Using a limited set of CT scans, we construct skull shape blendshapes that capture meaningful anatomical variation, and we further enhance the shape space using synthetic data generated by deep generative models. Finally, we train a regressor that maps FLAME facial parameters to skull parameters, enabling skull prediction from facial shape surface and resulting in a parametric skull model that is both easy to use and reproducible.

키워드