Automated Morphometric Analysis of Zebrafish Larvae for Developmental Toxicity Evaluation Using Deep Learning

Abstract

Accurate and automated analysis of zebrafish larval morphology is critical for developmental toxicity screening, especially within the context of New Approach Methodologies (NAMs). Here, we present a deep learning–based system designed to segment and quantify key anatomical features of zebrafish larvae, focusing on the eyes and swim bladder. Using fluorescence images acquired in a standardized ventral orientation, a DeepLab v3+ model with a ResNet-50 backbone was trained in MATLAB with customized data augmentation and a Tversky loss function to address class imbalance. Internal and external validation demonstrated high performance in terms of intersection-over-union (IoU), accuracy, and Dice coefficient, closely matching expert annotations. Our automated post-processing pipeline extracts biologically relevant parameters—including eye area, interocular distance, and swim bladder size—with high speed and reproducibility. Unlike previous approaches relying on lateral imaging, our method minimizes structural distortion and enables direct measurement of spatial relationships between bilateral organs. Segmentation overlays allow for visual verification and error detection. Although sensitivity to pigmentation or imaging artifacts remains a consideration, this system offers a scalable, transparent, and efficient solution for morphological evaluation. It sets the stage for further integration of cardiovascular and developmental endpoints using AI-powered image analysis.