AxoNet: A deep learning-based tool to count retinal ganglion cell axons

Area of Science:

• Ophthalmology
• Neuroscience
• Biomedical Engineering

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Accurate quantification of retinal ganglion cell axons is crucial for diagnosing and monitoring glaucoma.
• Existing automated methods for axon counting face limitations in accuracy and species adaptability.

Purpose of the Study:

• To develop a robust and extensible deep learning tool for automated axon counting in optic nerve (ON) images.
• To evaluate the performance of the developed tool against existing methods in both rat and non-human primate models.
• To establish a species-independent tool for analyzing ON tissue in neurodegenerative diseases.

Main Methods:

• Deep learning was adapted to regress pixelwise axon count density estimates.
• The tool, AxoNet, was trained on manually annotated rat ON images.
• AxoNet was evaluated on separate rat and non-human primate (NHP) ON datasets and compared to AxonMaster and AxonJ.

Main Results:

• AxoNet demonstrated superior performance compared to AxonMaster and AxonJ on both rat and NHP datasets.
• Performance was assessed using mean absolute error, R² values, and Bland-Altman analysis.
• The tool showed robustness to variations in tissue damage, image quality, and species.

Conclusions:

• AxoNet provides an accurate, automated, and extensible solution for counting retinal ganglion cell axons.
• The tool's species-independent nature makes it broadly applicable to glaucoma research and potentially other neurodegenerative diseases.
• AxoNet's deep learning approach eliminates reliance on hand-crafted features, enhancing its adaptability.