Multiscale learning and topological analysis across complex postures enable robust nematode size quantification in pharmacological assays

Abstract

Body size is an important trait that reflects animal development and physiology. In nematodes, precise measurement is valuable for linking variation in body dimensions to biological questions such as developmental timing, genetic regulation, and drug responses. However, robust size measurements can be difficult to obtain because nematodes can vary in curvature, have self-intersecting postures, and overlap with neighboring animals. Current image analysis software, such as CellProfiler, can measure isolated animals in straight postures but struggles with curly or overlapping animals. Here, we present NemaSize, an artificial intelligence (AI)-aided pipeline to measure Caenorhabditis nematode body sizes across complex postures using multi-scale learning and topology-aware skeletonization. Using You Only Look Once (YOLO) models trained at different spatial scales, NemaSize first identifies individual animals in a large field of view (FOV; 6.75 x 6.75 mm) and then performs high-resolution body segmentation in the region of interest (ROI). Next, NemaSize converts segmented body masks into topological graph representations, allowing curly and overlapping animals to be classified and skeletonized according to the body topology. NemaSize achieved less than 4% overall error in length and width measurements across all posture classes. Compared to CellProfiler, NemaSize demonstrated higher robustness for complex postures, including a 48% error reduction in length measurements for curly or self-overlapping animals. Application of NemaSize in high-throughput imaging assays further shows that NemaSize provides accurate quantification of Caenorhabditis briggsae larval development in response to the anthelmintic drug ivermectin, a task that was difficult for CellProfiler because of curly animal postures. Together, NemaSize provides a robust approach for automated body size quantification for Caenorhabditis nematodes and will support broad applications in high-throughput pharmacological and genetic screens. Beyond nematodes, NemaSize introduces a multiscale computational framework for analyzing elongated biological objects with complex topologies.