DrawerDissect: Whole-drawer insect imaging, segmentation, and transcription using AI

Elizabeth G Postema, Leah Briscoe, Chloe Harder, George R. A. Hancock, Lucy D. Guarnieri, Tony Eisel, Kelton Welch, Nicole Fischer, Christine Johnson, Diego Souza, Dexter Phillip, Rebekah Baquiran, Tatiana Sepulveda, Bruno A S de Medeiros 

1. Natural history museums often curate large collections of pinned insects. These collections represent invaluable records of biodiversity information, ecological patterns and phenotypic variation. A common goal of museums is to create digital versions of these records for curation and research purposes. However, traditional methods of specimen imaging and metadata transcription are prohibitively labor-intensive. High-throughput imaging of entire specimen drawers integrated with computer vision (CV) artificial intelligence (AI) models provides a potential solution.

2. Here we present DrawerDissect, a python-based pipeline for processing whole-drawer photographs, and a workflow to use it in entomological collections. By using custom CV models and large language models for text transcription, DrawerDissect can crop and segment specimens from images and extract metadata from specimen labels. DrawerDissect is flexible, customizable, and modular, allowing rapid downstream analyses of phenotypic features such as color, pattern, shape and size.

3. We used DrawerDissect to digitize the Field Museum’s (FMNH’s) entire tiger beetle (family Cicindelidae) collection, resulting in 13,484 high-resolution dorsal photographs, masked specimen images, and basic body measurements. All specimens are linked to taxonomic and biogeographic data. We also extracted specific location metadata for 3,648 specimens. We then provide an example of using DrawerDissect outputs with existing color analysis methods in ImageJ to investigate taxonomic and geographic differences in coloration. Finally, we trained an accurate species identification model, Cicindel-ID, using ~7,000 masked images of specimens in the genus Cicindela.

4. DrawerDissect’s novel multi-model AI workflow provides an efficient and reproducible framework that meets the demands of high-throughput digitization of natural history museum collections, unlocking the research potential of large specimen collections.