Abstract
INTRODUCTION: Conventional computational modeling of human body and tissue relies on time-consuming reconstruction from medical images. Instead, generative artificial intelligence synthesizes novel samples with few marginal costs. However, existing models are incompatible with 3D data structures. This study aims to establish a 3D generative tissue modeling pipeline. The key objective is to address the following technical bottlenecks: 1) tackle the neural-network incompatibility of discretization-agnostic meshes, 2) formulate a sequence-type architecture for shape prior learning and generation, and 3) realize smooth source-to-target mapping. METHODS: We propose a data-driven non-rigid registration that learns geometry-informed features through self-supervised pretraining, and infers high-fidelity correspondences accordingly. A template mesh is uniformly registered onto 90 femur shapes to align inter-vertex connectivity. A variational autoencoder (VAE) is trained on the structurally aligned samples, and synthesizes novel shapes thereafter. Left femur of a baseline finite-element model is mapped onto the synthesized instances respectively, and the femurs are computationally loaded to investigate biomechanics with diverse morphologies. RESULTS: Tuning the six latent dimensions of the trained VAE independently generates shapes in heterogenous morphological patterns, i.e., overall size, overall bending curvature, slenderness, shaft length, fine-level shaft curvature, and local epiphysis/metaphysis style. Quantitatively, varying the first latent results in an 83.4 mm change in femur length, while the second latent controls the equivalent radius of shaft segment in a range from 66.8 mm to 862.6 mm. The VAE model synthesizes geometrically valid shapes within up to 3 standard deviations (>99.7%) of the entire latent. 10 new femurs are synthesized and registered onto a baseline finite-element model in less than 100 s for each case. Preliminary analysis with three-point bending load reveals morphological variation might have a significant influence on deformation pattern and bending biomechanics. CONCLUSION: This study establishes a contemporary generative paradigm for tissue modeling, and demonstrates efficacy and feasibility of biomechanics investigation with synthetic shapes. Our method produces high-fidelity, simulation-ready models in only minutes. The pretraining scheme is scalable to multiple anatomical structures and sheds light to foundation model of 3D anatomies, which might promisingly benefit a lot of production workflows, e.g., active-passive vehicle safety, robot-assisted surgery, and all biomechanics/morphology-relevant tasks.