June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
AxonClassNet: A Deep Learning Approach for Automated Segmentation and Grading of Retinal Ganglion Cell Axons
Author Affiliations & Notes
  • Shelby Graham
    The University of Tennessee Health Science Center, Memphis, Tennessee, United States
    The University of Memphis, Memphis, Tennessee, United States
  • Kyle Freeman
    The University of Tennessee Health Science Center, Memphis, Tennessee, United States
  • Sophie Pilkinton
    The University of Tennessee Health Science Center, Memphis, Tennessee, United States
  • Madhusudhanan Balasubramanian
    The University of Memphis, Memphis, Tennessee, United States
  • Monica M Jablonski
    The University of Tennessee Health Science Center, Memphis, Tennessee, United States
  • Footnotes
    Commercial Relationships   Shelby Graham None; Kyle Freeman None; Sophie Pilkinton None; Madhusudhanan Balasubramanian None; Monica Jablonski None
  • Footnotes
    Support  NEI EY001200 and RPB Challenge Grant
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 389. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Shelby Graham, Kyle Freeman, Sophie Pilkinton, Madhusudhanan Balasubramanian, Monica M Jablonski; AxonClassNet: A Deep Learning Approach for Automated Segmentation and Grading of Retinal Ganglion Cell Axons. Invest. Ophthalmol. Vis. Sci. 2023;64(8):389.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Retinal ganglion cell (RGCs) necrosis and loss are hallmarks of degenerative retinal diseases such as glaucoma. Characteristics of axonal distribution in the optic nerve (ON) such as axon count, axon morphology and the state of necrosis of each axon are useful for staging the disease and its progression. While a few computational methods are available for detecting axons from ON images, no methods are available at present for automatic markup and grading of ON axons. We present a deep learning approach for ON instance detection capable of counting, segmenting, and labeling axons as live or necrotic.

Methods : ONs harvested from several generations of outbred heterogenous stock NIH rats were prepared, embedded, and stained with p-phenylenediamine. In an initial set of 13 confocal images from 1 rat, all ON axons were manually annotated as live or necrotic by 3 trained reviewers. Any discrepancies in annotations were reviewed and adjudicated by an expert reviewer. A region based convolutional neural network (Mask RCNN) architecture was initialized with pretrained ImageNet weights and was fine-tuned using annotated ON images for detecting and grading ON axons (AxonClassNet). An initial model was identified using the initial 13 images (9 training; 2 validation). Using this model, a separate set of 74 images from 4 rats were automatically annotated and then reviewed and corrected by 3 trained reviewers. A final model was identified from the second phase of training the AxonClassNet using a larger set of images (25 training; 12 validation). Performance of the AxonClassNet was assessed using its precision, recall and F1-score in correctly detecting and grading ON axons.

Results : The average precision, recall and F1-score for correctly segmenting and grading necrotic axons were 86.81%, 93.91% and 90.17% respectively. For segmenting and grading live axons, the average precision, recall and F1-score were 79.06%, 92.91% and 85.41% respectively.

Conclusions : AxonClassNet segmented all ON healthy and necrotic axons in confocal images with high accuracy. A higher recall measure with a relative lower precision indicated that the model identified nearly all axons but with a few grading errors. With additional training, our deep learning approach shows promise for high throughput localization and grading of ON axons in various experimental glaucoma studies and assessing ON health with candidate drugs.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×