Automated Identification of Histological Lesions in Nonmodel Organisms: Reinvigorating Environmental Science

Abstract

In response to contaminants, alterations of tissue morphology and function can reflect ecosystem health and be irreversible. Traditional assessments require highly trained personnel, are cost- and time-ineffective and produce subjective, qualitative results susceptible to bias. Automated digital histology aims to address these challenges while relieving the burden on pathologists and increasing the meaningfulness and reproducibility of findings. Over the past few years, technological advances in image recognition facilitated the analysis of content-dense histological images for human health. Herein, we applied such advancements to environmental science, and automated digital histology was tested for the identification of lesions (steatosis, melano-macrophage aggregates, leucocyte infiltration, and granuloma) in livers from two fish species (Gadus morhua and Limanda limanda) sampled near oil and gas platforms. Images scored by professional histopathologists were used to train a machine learning model on hematoxylin and eosin-stained whole slide images. The automated digital detections corresponded well with traditional assessments (between 85% and 95%) but required less expertise and were faster. Our results demonstrate a viable path toward a complete, automated workflow for fast, cost-effective and accurate digital histology analysis of environmental and ecotoxicology samples. Efficient, accessible digital histology models will advance understanding of ongoing environmental changes and enhance future preparedness.