From images to evolutionary traits: Automated morphological representation of insect phenotypes

Date: May 19th(Tue) 
Time: 14:00 - 15:45
Room: C210 

Title: From images to evolutionary traits: Automated morphological representation of insect phenotypes 

Presenter:  Mr. Ancheng Peng from Muséum national d'histoire naturelle

Abstract:
Molecular data have expanded rapidly in recent decades, yet morphology remains essential for evolutionary biology, especially for integrating fossils and linking phenotypic variation to ecological and functional change. However, quantifying morphology at evolutionary scale is still constrained by manual landmarking, subjective character definition, and the difficulty of comparing complex structures across broad taxonomic ranges. We developed an automated framework designed to extract morphological representations from image-based phenotypic datasets while reducing dependence on manual annotation. The workflow uses a self-supervised residual attention autoencoder to extract morphological representations, with attention maps used to visualize the anatomical regions contributing to the learned features. We first tested the framework on 2D insect wing datasets, where the extracted features facilitated downstream evolutionary analyses. Small-scale tests on three-dimensional datasets further suggest that the framework can be extended beyond 2D images, providing a basis for future applications in 3D phenomics. This approach aims to move morphology from manually curated descriptors toward scalable, reproducible, and interpretable phenotypic datasets for evolutionary biology.