SegBio: A lightweight end-to-end toolkit for Instance Segmentation of biological samples

Abstract

High-throughput phenotyping of biological samples is essential for large-scale studies but is frequently bottlenecked by the need for accurate instance segmentation in crowded images. While deep learning offers powerful solutions, the high cost of manual annotation and the requirement for coding expertise often limit adoption in routine laboratory workflows. Here we present SegBio, a lightweight, open-source pipeline that enables end-to-end instance segmentation for non-expert users. The protocol features an interactive annotation GUI that extrapolates full masks from minimal centerline markings, significantly reducing manual labeling effort. It further integrates a configurable U-Net training module and a standalone inference application with a human-in-the-loop editing workflow for rapid and intuitive error correction. We employ the pipeline to annotate and train the model on a novel dataset of crowded C. elegans images. Validated on independent datasets, SegBio achieves high segmentation performance (Panoptic quality [~]0.85) and accurately quantifies per-animal morphology and fluorescence. By eliminating external dependencies and streamlining the correction process, SegBio provides a scalable solution for routine phenotyping that is easily generalized to other crowded biological samples, such as cellular organelles, cells, and organisms.