Purchase this article with an account.
Bailey Hannon, Matthew Ritch, Robert Christian Ford, Arthur Thomas Read, Kristin Gao, Eva L. Dyer, Juan Reynaud, Grant Cull, Claude F Burgoyne, Machelle T Pardue, C Ross Ethier; Machine Learning-Based Quantification of Axonal Damage in Glaucomatous Rat Optic Nerves. Invest. Ophthalmol. Vis. Sci. 2019;60(9):2854.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Glaucoma is characterized by progressive degradation of retinal ganglion cell axons in the optic nerve (ON). Glaucomatous damage is often quantified by manually or automatically counting healthy ON axons. However, quantification of other axonal features (e.g. myelin sheath thickness and axon size distributions) would be useful in describing glaucomatous damage patterns. Towards this goal, we have utilized machine learning software to count healthy axons in rat ONs and validated results vs. hand counts and existing software.
We used an open-source machine learning platform, ilastik, to analyze images of rat ONs with a range of axonal loss. First, 6 subregions of rat ON histological sections were used to train ilastik’s pixel segmentation and object classification modules. Once trained, 25 “calibration” subregions in which axons had been manually counted by 4 trained personnel were automatically counted by ilastik, and an iterative process of re-training the software was used to adjust ilastik to best match the manual counts. Finally, 25 additional “validation” subregions were manually counted, and then automatically counted by ilastik and previously published automated software, AxonMaster9 (AM9; Reynaud et al. 2012). A Deming regression was fit to both the calibration and validation data sets to adjust for systematic errors and the resulting correction equation was used to match manual and automated counts.
Automated counts correlated well with manual counts for all ON images over a range of axon loss. The mean absolute value error was 14% of the total axons in the image and residuals were randomly dispersed around zero (Fig. 1A). Further, automated counts on the validation set of images for ilastik and AM9 had mean absolute errors within 10% of manual counts (Fig. 1B).
Our ilastik-based automated axon counting software was valid for basic axon counting, with acceptably small error and bias. After correction equations were applied in the validation set of images, ilastik estimated manual counts with similar accuracy to existing AM9 software. Future work will include continued validation and quantification of axon size and myelin sheath thickness utilizing the object classification capabilities of ilastik.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Fig. 1: A. Calibration of ilastik (red) vs. manual counts and B. validation of ilastik and AM9 (purple) vs. manual counts with Deming regression fits (n=25 images/dataset).
This PDF is available to Subscribers Only