Deep learning-based counting and body-area measurement enable sensitive population-level endpoints in the Collembola reproduction test

Abstract

Springtails (Collembola) are standard test organisms for soil ecotoxicological assessment; under the ISO 11267 and OECD 232 guidelines, adult abundance and juvenile production are the primary endpoints. However, these rely on labor-intensive manual counting and capture only abundance, overlooking sublethal effects on growth. We developed an automated system that couples YOLOv8 object detection with Voronoi-Otsu segmentation-based body-area measurement, so that adults and juveniles of Allonychiurus kimi are counted and their growth (body area) measured simultaneously, and applied it to toxicity assessment based on average and cumulative body area and to cohort analysis of juvenile body-area distributions using a Gaussian mixture model (GMM). The model was trained on 142 annotated images and validated on 160 images from toxicity tests with cadmium (Cd), arsenic (As), and copper (Cu), and both detection accuracy and agreement with manual measurements were excellent (coefficient of determination, R2 >= 0.95). Cumulative body area, a composite indicator integrating abundance and individual size and thus reflecting both reproduction and growth, showed the lowest EC50 in all three metal(loid)s (Cd 14.10, As 10.53, Cu 398.80 mg kg-1). Cohort analysis showed that the two cohorts present in the controls merged into one at the highest concentrations (As 20, Cu 800 mg kg-1), indicating that toxic stress drives not only individual-level suppression of reproduction and growth but also population restructuring. Beyond accurately and efficiently replacing manual counting, the system adds more sensitive composite endpoints and population-level analysis, providing a more comprehensive, scalable framework for soil ecotoxicological risk assessment.