Machine learning‑based segmentation and classification of axons following optic nerve injury — The Association Specialists
  1. Schedule
  2. Poster Presentations
  3. Machine learning‑based segmentation and classification of axons following optic nerve injury

Machine learning‑based segmentation and classification of axons following optic nerve injury (22036)

Parth Patel 1 , Sarah Hellewell 1 2 3 , Brittney Lins 1 2 , Terry McGonigle 1 , Andrew Warnock 1 , Naing Lin 1 , Enoch Teo 1 , Carole Bartlett 1 , Melinda Fitzgerald 1 2 , Chidozie Anyaegbu 1 2
  1. Curtin University, Canning Vale, WA, Australia
  2. Perron Institute for Neurological and Translational Sciences, Sarich Neuroscience Research Institute, Nedlands, Western Australia, Australia
  3. Centre for Neuromuscular and Neurological Disorders, University of Western Australia, Crawley, Western Australia, Australia

Changes to axonal ultrastructure provide important pathological insight into secondary degeneration following manual and semi‑automated methods for quantifying axonal morphometrics require significant user input, making them labour-intensive. AxonDeepSeg is a convolutional neural network for automatic morphometric feature analysis of micrographs. We hypothesised that AxonDeepSeg could be used to derive both traditional and new axon and myelin metrics while retaining high fidelity to manual measurements. We trained an AxonDeepSeg model to segment axons in transmission electron micrographs of optic nerves from female Piebald Virol‑Glaxo rats following a partial transection injury (n=5) or sham surgery (n=5). Our model segmented 68 images with a 94.7% pixel-wise accuracy and was comparable to manual measurement of axon diameter. Using AxonDeepSeg measurements, we found that injury increased the g-ratio (p<0.03), solidity (p<0.02), and axon area/fibre area ratio (p<0.02). Unsupervised spectral clustering revealed several clusters of axons with unique defining metrics. Injury significantly altered these cluster metrics, resulting in a larger g-ratio (p=0.02), solidity (p=0.03) and axon area/fibre area ratio (p=0.03) in cluster 1 and larger g-ratio (p=0.02), axon area (p<0.001), axon diameter (p=0.03), solidity (p=0.02), and axon area/fibre area ratio (p=0.02) in cluster 2. We conclude that AxonDeepSeg accurately reproduced manual measurements of axon morphology, and enabled discovery of novel morphometrics that distinguish injured axons from uninjured axons, allowing charting with injury. AxonDeepSeg will be a valuable tool for objective and high-throughput analysis of axonal morphology, facilitating future investigations in neurotrauma and neurodegeneration.