Researchers from Sciome and the National Institute of Environmental Health Sciences (NIEHS) have developed an advanced artificial intelligence (AI) system that could transform how scientists screen chemicals for developmental toxicity.
The research, conducted as part of the Systematic Evaluation of the Application of Zebrafish in Toxicology (SEAZIT) project, introduces a deep learning approach to analyzing images of zebrafish embryos—an established model for studying the effects of chemical exposures during early development. Traditionally, trained scientists must manually examine these images to detect morphological changes in the embryos, a process that is both time-consuming and vulnerable to human bias.
Now, thanks to this joint effort, deep learning models can automate and standardize this task. By training AI systems on a large set of labeled images from the SEAZIT project, the team developed powerful classification and segmentation tools capable of identifying and measuring key developmental features in the embryos.
One of the models, a multi-view convolutional neural network (MVCNN), demonstrated impressive performance—correctly distinguishing between normal and affected embryos with an F1 score of 0.88. Additional models targeting specific developmental abnormalities achieved similarly strong results, especially when related abnormalities were grouped.
The team also built segmentation models to precisely map and measure specific regions of the embryos. Most of these models achieved a high accuracy rate, with 9 out of 11 reaching an Intersection over Union (IoU) score of 0.80 or better.
This innovative work marks a significant step forward in the field of toxicology, offering a faster, more objective, and highly reproducible method for screening potential chemical hazards. The collaboration between Sciome and NIEHS highlights the power of combining cutting-edge technology with public health research to improve chemical safety testing.
Read the original article here.
