From traditional histopathology to virtual staining: deep learning approaches for assessing tissue lesions in farmed fish.

Pesticides in aquaculture water and feed primarily induce chronic, sublethal toxicity rather than acute mortality, impairing growth, immune function, and tissue integrity. Consequently, organ-specific histopathological biomarkers are essential for ecotoxicological assessment and welfare monitoring. Traditional histopathology directly evaluates structural damage in key organs (e.g., gills, liver, kidneys). However, manual workflows face major bottlenecks, including invasive sampling, labor-intensive preparation, and observer-dependent scoring that hinder high-throughput screening and regulatory harmonization. This review evaluates how digital pathology integrating Whole Slide Imaging (WSI), deep learning for automated lesion quantification, and generative AI-based virtual staining bridges these analytical gaps. We examine how digitalization enhances diagnostic objectivity and scalability while reducing scoring variability. Furthermore, we critically address key implementation challenges, including ground-truth requirements, domain shifts across staining protocols, species generalization, and dataset annotation burdens. Ultimately, AI-enabled digital pathology represents a credible pathway toward objective, high-throughput pesticide toxicity assessment in aquaculture. However, standardized validation frameworks and robust ground-truth datasets remain critical prerequisites before these tools can transition from experimental promise to regulatory acceptance.