Deep Learning Image Analysis of High-Throughput Toxicology Assay Images.

Area of Science:

• Cell biology
• Computational biology
• Toxicology

Background:

• High-throughput chemical screening generates vast image data requiring manual analysis.
• Automating image classification is crucial for efficient drug discovery and toxicity testing.

Purpose of the Study:

• Develop and validate a deep learning method for automated classification of cell morphology in assay images.
• Assess the accuracy and reproducibility of deep learning models in identifying chemical-induced cytotoxicity.

Main Methods:

• Trained convolutional neural networks (CNNs) for binary (healthy/altered) and multi-class (healthy/intermediate/altered) image classification.
• Utilized Class Activation Maps (CAM) to visualize and understand CNN decision-making processes.
• Tested various CNN architectures (ResNet 34, 50, 101) on diverse cell culture datasets (primary hepatocytes, HepaRG cells).

Main Results:

• Binary CNN classifier achieved >98% accuracy; multi-class classifier achieved >95% accuracy.
• Demonstrated strong correlation between chemical dosage and classifier-predicted scores.
• Identified specific image regions used by CNNs for classification via CAM, enabling model optimization.

Conclusions:

• Deep learning provides a highly accurate and reproducible method for automated cytotoxicity assessment.
• Automated image classification significantly reduces manual analysis time in chemical screening.
• The developed method is applicable across various cell types and can aid in characterizing dose-response relationships.