A Foundational Model for Biomedical Shapes

Geometric representations of anatomical structures, specifically meshes and point clouds, are being actively researched in medical imaging for applications such as developing neural surrogates, statistical shape modeling, brain parcellation, and biomarker detection. In the medical domain, where high-quality labeled data is notoriously difficult to acquire, these representations are highly sought after because they provide a compact, geometric alternative to high-dimensional voxel data, effectively reducing noise in data-scarce environments.

In recent years, foundational models have revolutionized representation learning in the language and vision domains by leveraging large-scale datasets to achieve superior generalization, leading to widespread commercial adoption. These models offer a robust solution for downstream applications with limited training data. This project aims to bridge these fields by developing a foundational model specifically tailored to organ geometries, with downstream applications in biomechanics surrogate models. By training a shape encoder on a large set of shapes, we aim to provide a versatile framework for medical applications currently constrained by data sparsity and the complexities of anatomical variation.