Related Experiment Video
Updated: Jun 27, 2026
08:57
Quantification of Optic Nerve Cross Sectional Area on MRI: A Novel Protocol using Fiji Software
Published on: September 4, 2021
Comparison of Deep Learning Tools for Optic Nerve Axon Quantification Finds Limited Generalizability Upon Independent
Benton Chuter1, Noah Emmert2, Min Young Kim1
1Department of Ophthalmology, Hamilton Eye Institute, University of Tennessee Health Science Center, 930 Madison Avenue, Memphis, TN 38163, USA.
Bioengineering (Basel, Switzerland)
|June 26, 2026
Summary
Machine learning models for optic nerve histology show high performance within studies but struggle to generalize to new datasets. Further validation is needed for reliable automated axonal injury assessment in glaucoma research.
Area of Science:
- Ophthalmology
- Computational Biology
- Histology
Background:
- Machine learning (ML) offers objective quantification of optic nerve histology for assessing axonal injury in experimental glaucoma.
- The generalizability of existing ML models to independent datasets is not well-established.
Purpose of the Study:
- To evaluate the performance and generalizability of publicly available machine learning models for automated optic nerve histology quantification on novel datasets.
Main Methods:
- A scoping review following PRISMA-ScR guidelines identified three deep learning models: AxoNet, AxonDeepSeg, and AxoNet 2.0.
- These models were independently validated on rat and mouse optic nerve datasets.
- Performance metrics included correlation coefficients, precision, recall, and Dice coefficients.
Main Results:
- Models demonstrated high performance (correlation coefficients 0.959-0.99) in original studies.
- Validation on independent datasets showed reduced performance: correlation (r=0.831-0.907 in rats, r=0.57-0.74 in mice), low recall (0.18-0.27 in rats), and Dice coefficients (0.29-0.40 in rats).
- Model rankings varied across datasets, indicating domain shift and scale-dependent effects.
Conclusions:
- Publicly available machine learning models for optic nerve histology exhibit strong within-study performance but limited generalizability to independent datasets.
- Standardized validation datasets and multi-center testing are crucial for improving the reliability of these tools in glaucoma research.
