This paper introduces Morpho-VAE, a landmark-free deep learning framework for shape analysis using image data. Morpho-VAE, a modified variational autoencoder (VAE), combines unsupervised and supervised learning to extract morphological features distinguishing different labeled classes. Applied to primate mandible images, it effectively clustered samples by family, even without correlation to phylogenetic distance. It also demonstrated reconstruction of missing segments in incomplete images, making it a versatile tool for biological shape analysis.

Masato Tsutsumi, Nen Saito, Daisuke Koyabu, Chikara Furusawa