LCSM: A Local Chart-Based Statistical Model for Femoral Shape Analysis

Abstract

Statistical shape modeling of the femur is vital for clinical applications, yet linear methods struggle with complex geometric variations. While recent chart-based generative models like ChartPointFlow excel at general point cloud reconstruction, their direct application to the nuanced domain of SSM is limited. To bridge this gap, we propose a novel framework that adapts and significantly extends the chart-based paradigm for robust statistical shape analysis. Our core contributions are threefold. First, we are applying this substructure-based generative approach to the statistical modeling of anatomical shapes, moving beyond reconstruction. Second, to overcome the limitations of baseline encoders in capturing intricate anatomical details, we design a Geometric-enhanced Dynamic Graph Convolutional encoder that integrates multi-level geometric feature encoding. Third, to better learn the statistical distribution across a population, we introduce a point-wise feature augmentation module, termed Fine Feature Mixture module, which enriches feature diversity and robustness by fusing fine-grained representations across different samples. Extensive reconstruction and generation experiments on a femur dataset demonstrate our framework’s superiority. Furthermore, evaluations on multiple ShapeNet categories validate its generalization and applicability to diverse 3D shape modeling tasks. Code will be available in https://github.com/LujZhang/LCSM