U.S. flag

An official website of the United States government

Dot gov

The .gov means it's official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you're on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Artificial Intelligence (AI)-Driven Morphological Assessment of Zebrafish Larvae for Developmental Toxicity Chemical Screening

Arpit Tandon1, Brian E. Howard1, Adrian J. Green1, Rebecca Elmore1, Ruchir Shah1*, Alex Merrick2, Keith Shockley3, Kristen Ryan2, Jui-Hua Hsieh2

1Sciome LLC, Research Triangle Park, NC, United States
2Division of Translational Toxicology, National Institute of Environmental Health Sciences (NIEHS), Research Triangle Park, NC, United States
3Division of Intramural Research, National Institute of Environmental Health Sciences (NIEHS), Research Triangle Park, NC, United States

DOI: https://doi.org/10.22427/NTP-DATA-500-016-001-000-3
Publication DOI: https://doi.org/10.1016/j.aquatox.2025.107415

Publication

Abstract

Screening chemicals using the zebrafish embryo developmental toxicity assay requires visual assessment of larval morphological changes based on images by experienced screeners. The process is time-consuming and prone to subjectivity. However, deep learning models trained using labeled image data offer a more objective and efficient alternative. As part of the Systematic Evaluation of the Application of Zebrafish in Toxicology (SEAZIT) project, which aims to provide a scientific basis for programmatic decisions on the routine use of zebrafish in toxicological evaluations, we developed deep-learning classification and segmentation models to support image-based assessments. Using labeled SEAZIT image data from embryos exposed to various chemicals for 5 days, we trained classification models to detect 20 distinct types of larval morphological changes. We also developed segmentation models to identify and measure larval regions of interest. In a baseline binary classification task distinguishing normal embryos from those with any morphological change, our multi-view convolutional neural network (MVCNN) achieved an F1 score of 0.88, demonstrating strong overall performance. Among classifiers for specific morphological changes, eight achieved F1 scores above 0.70. Grouping related abnormalities further improved performance, with five out of seven grouped classifiers reaching F1 scores near 0.80. Segmentation models for most of the regions of interest (9 out of 11) have an Intersection over Union (IoU) score of at least 0.80. Together, these models facilitate rapid, standardized and reliable evaluation of larval morphological changes in zebrafish images, reducing the need for subjective manual review and enhancing reproducibility in development toxicity screening.

Methods in Brief

To enhance toxicological screening efficiency, Sciome LLC implemented novel deep learning classification models, customized for the SEAZIT (Systematic Evaluation of the Application of Zebrafish in Toxicology) program capable of classifying 20 different types of zebrafish larval morphological changes. These models were trained using labeled image data from larvae exposed to various chemicals over five days. In addition to classification models, deep learning segmentation models were also deployed to identify and quantify key larval regions of interest, such as the eyes, bladder, yolk sac, notochord, etc.

The image data used for training and validation were sourced from zebrafish larval images collected during the two phases of an interlaboratory study (SEAZIT): the Dose Range-Finding (DRF) phase and the Definitive (DF) phase.

Conclusion

The classification model shows high accuracy for several of the most common morphological changes evaluated in a chemical screening approach and could allow researchers to quickly screen the bulk of zebrafish larva images exposed to different chemicals, while allowing expert screeners to focus on the more ambiguous or unusual images. Segmentation models also show high accuracy for most of the regions and will allow users to quantify the effects of a chemical exposure more accurately and to perform a detailed downstream analysis of the resulting morphological changes. As a result, these automated models can save time while producing results that are more accurate, more reproducible, and less biased than manual assessments.

Data Release Information

The image data was collected during the two phases of an interlaboratory study (SEAZIT): Dose Range Finding (DRF) phase and Definitive (DF) phase. Images indicated in Metadata_DRF_DF.xlsx were used in the training and validation of the deep learning models. The images are saved in folders according to the phase and plate type. The Annotation sheet in Metadata_DRF_DF.xlsx contains annotations for each of the image files along with study source, chemical and dose information. The Ontology sheet in Metadata_DRF_DF.xlsx provides ontology mapping for the annotations based on the Zebrafish Phenotype Ontology. The code used for this analysis can be available upon request (software.support@sciome.com)

Data

Image Data