Vert-NSM: neural shape models for shape completion, classification, and landmark-free morphometrics of vertebrate specimens

Katherine Aubry Wolcott , Amar Cutura, Matthew Delage, Kelsey Fenner, Ali Sean Khan, Jia Xuan Koh, Mitchell Riegler, Colin Sawyer, Joshua Yu, Anthony Gatti, Edward Stanley, Arthur Porto 

Natural history museums are building digital 3D archives at a scale that enables geometric morphometric analysis across thousands of specimens. Landmark-based analyses depend on anatomical priors and often exclude unidentified and incomplete specimens, which are abundant in collections and the fossil record. Continuous shape representations offer a landmark-free alternative that can be combined with neural networks to perform shape completion and classification. It is unknown how their morphospaces compare with landmarks or how accurately they perform. We present Vert-NSM, neural shape models for vertebrate biodiversity specimens, and apply them to squamate vertebrae, a well-represented, but challenging test case. Models were trained on 2,896 vertebral meshes from 95 specimens across 40 families, produced in part by the oVert digitization initiative. We evaluated shape completion, classification, and morphometrics of sparse landmarks, dense correspondences, and Vert-NSM latents. Shape completion accuracy reached 0.375% of a vertebral radius. Classification exceeded frequency-weighted chance at nearly every level tested, reaching 80% accuracy for the cervical spine and 90% for Cordylidae. Vert-NSM shared the dominant axes of variation recovered by landmarks, and identified shape differences that did not depend on anatomical priors. In downsampling experiments, models performed well when trained with at least 30 specimens. Vert-NSM has the potential to unlock discoveries from vertebrate collections at scale and in deep time.