Developing a deep learning-based image analysis model for high-throughput micronucleus assays: Genotoxicity as a sediment quality indicator in East Taihu and Yangcheng Lakes, China

Sediment quality has become a growing concern, as sediment-bound anthropogenic pollutants, particularly genotoxic compounds, may serve as secondary pollution sources, posing significant risks to aquatic ecosystems and human health. The in vitro micronucleus (MN) assay, standardized by ISO and OECD guidelines, is widely used for genotoxicity assessment; however, traditional manual MN scoring is labor-intensive, time-consuming, and susceptible to observer bias. Therefore, this study proposed a deep learning-based model for automated identification and quantification of cell nuclei and MN. Among the three trained models, the architecture incorporating hierarchical attention mechanisms, including self-attention and channel-spatial attention, was selected due to its superior segmentation performance. Compared with manual scoring, the model showed 95.63% (nuclei) and 97.38% (MN) agreement in Bland-Altman analysis, while achieving processing speeds approximately 20-fold higher per hour and 60-fold higher per day. Using the model, sediment genotoxicity from East Taihu and Yangcheng Lakes, two major freshwater systems and drinking water resources in China heavily impacted by human activities, was evaluated under both rat-S9 metabolically active and inactive conditions. Significant genotoxicity was observed, with minor discrepancies between manual and model counts, primarily in weakly genotoxic samples. Genotoxicity decreased following S9-activation, likely due to metabolic detoxification or inhibitory effects of co-existing substances. Regardless of metabolic activation, Yangcheng Lake sediments consistently exhibited higher genotoxic effects than those from East Taihu Lake. As a proof-of-concept application of deep learning in environmental genotoxicity assessment, the model architecture has been made publicly available to support high-throughput MN assay applications.