Georgia Institute of Technology
Deep Learning-Based Analysis of Optic Nerve Axonal Morphology
Abstract
dc:description.abstractAxon loss and degeneration are used to quantify the progression of several neurodegenerative diseases, such as glaucoma and multiple sclerosis, in animal models. In glaucoma, the gold standard for quantifying optic nerve health post-mortem is manual counting of axons from micrographs, which is subjective and laborious. The present research is focused on developing a deep learning model to identify and segment normal-appearing axons, including their axoplasm and myelin sheath, from light micrographic montages of whole optic nerve cross-sections, AxoNet 2.0. These segmentation maps are fed into an image-processing pipeline for post-processing and for computing morphometric properties such as axoplasmic area, eccentricity, diameter, etc. The morphometric parameters of RGC axons segmented by our software agreed well with manual annotations for validation and testing sets as quantified by high values of soft-Dice coefficients (0.81 ± 0.02), low mean absolute percentage errors (< 15%) for the computation of various morphological properties, and high sensitivity (0.92) and precision (0.83). This software also showed improved axon counting accuracy compared to our previously-developed software, AxoNet 1.0, as shown by higher R2 values obtained from regressing automated counts of normal-appearing axons vs. ground truth (manual) axon counts in training (0.98 vs. 0.92), validation (0.90 vs. 0.84) and test (0.92 vs. 0.85) datasets. AxoNet 2.0 also performed well on optic nerve cross-section images collected from other animal models (R2 between automated vs. manual counts = 0.97 for mice and 0.98 for non-human primates). AxoNet 2.0’s performance was also compared to AxonDeep, a deep learning algorithm trained to identify only axoplasm in RGC axons from light micrographs of mouse optic nerve cross-sections. AxoNet 2.0 showed similar performance to AxonDeep ((R2 = 0.98 for AxoNet 2.0 counts vs. manual counts; R2 = 0.97 for AxonDeep vs. manual counts). Applying this automated software to images of previously collected normotensive optic nerves, we created a reference retinal ganglion cell axonal atlas for Brown-Norway rats, an animal model widely used in glaucoma research. We observed that RGC axon counts and morphometrics in the atlas agreed well with previous reports on manual analysis of optic nerve axons and that morphological parameters exhibited a wide range of coefficients of variation; interestingly, the coefficient of variation for myelin sheath morphometry was only ~2%. Hence, we developed damage metrics based on myelin sheath morphometry. These metrics strongly correlated with an independent damage score (p<0.001). Additionally, we developed a metric based on preferential loss of large axons in hypertensive vs. normotensive eyes. However, this latter metric correlated less well with the independent damage score. This technology can be used to objectively evaluate changes in RGCs and overall levels of damage while standardizing RGC axon assessment in future glaucoma studies. AxoNet 2.0 and our reference atlas could help us answer important questions such as “How does axon morphology change with increase optic nerve damage?”, and “Which optic nerve regions are protected by various treatments?”.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Goyal, Vidisha
- Advisor dc:contributor.advisor
-
- Ethier, C. Ross
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/72441