Abstract
In-depth exploration of the effects of genes on the development, physiology, and behavior of organisms requires high-precision phenotypic analysis. However, the overlap of body postures in group behavior and the similarity of movement patterns between strains pose challenges to accuracy analysis. To address this issue, we designed the WormYOLO model based on the YOLO architecture, which improves the segmentation performance of C .elegans and effectively handles overlapping poses in images. In detection and segmentation tasks, WormYOLO performs well on the more overlapping Mating dataset, with its object detection performance improving by 24.1% (mAP(0.5:0.95)) compared to Deep-worm-tracker, and its segmentation performance improving by 9.3% (mAP(0.5:0.95)) compared to WormSwin. In addition, we propose a more accurate novel bending counting algorithm. In experiments, WormYOLO segmented images, followed by a feature point extraction algorithm to identify changes in worm skeleton positions, ultimately quantifying behavioral features with a counting algorithm. We conducted analytical experiments on various mutant strains based on their motion characteristics, investigating behavioral differences among the strains and assessing the correlation between high-dimensional phenotypic traits and relative lifespan.