AxoNet 2.0: A deep learning (DL)-based tool for morphometric analysis of retinal ganglion cell (RGC) axons

Abstract

Purpose : A key outcome measure in animal models of glaucoma is the number and appearance of RGC axons, typically obtained from light micrographs of optic nerve (ON) cross-sections. However, this analysis is time-consuming and subjective. Here we expand on our RGC axon counting software (AxoNet 1.0, [1]) to also identify axoplasm (“Ax”) and myelin sheath (“MySh”) of normal-appearing RGC axons and extract their morphometric features.

Methods : Micrographs from 12 control and 14 hypertensive rat ONs were acquired as before [1], representing a wide range of glaucomatous damage. 1421 12x12 um² sub-images were randomly selected from these micrographs, and Ax and MySh of normal-appearing axons were annotated by 16 trained individuals with cross-validation. Sub-images were randomly divided into training (90%), validation (5%), and test (5%) sets. A trained U-Net architecture generated segmentation of Ax and MySh of the normal-appearing axons, assigning a probability of each pixel being Ax or MySh. These maps were processed through an image processing pipeline to compute, for each axon: area, convex area, perimeter, eccentricity, average MySh probability (average probability over all MySh pixels of each pixel being MySh), and average Ax probability. Outcome measures included: (i) the soft-dice coefficient between predicted and ground-truth segmentation maps; (ii) R² regressing number of the normal-appearing axons counted automatically vs. ground truth. [1] Ritch+, Sci Rep, 2020

Results : The model performed well (Fig 1) with high soft-dice coefficients (Fig 2). AxoNet 2.0 outperformed AxoNet 1.0 as judged by agreement with ground truth (Fig 2).

Conclusions : A DL model can segment Ax and MySh of normal-appearing axons from ON sub-images with varying ON health. This approach will speed RGC axon counting and morphometric analysis in animal models of glaucomatous optic neuropathy.

This is a 2021 ARVO Annual Meeting abstract.

 

AxoNet 2.0 performance on a representative image. (A) Original sub-image (B) Annotated image (pink=Ax, yellow=MySh) (C) Predicted Ax probability map (greyscale bar: Ax probability) (D,E) Ax and MySh of segmented axons, respectively (F) Features computed for a segmented axon.

 

(A) Average mean soft-dice coefficients, a measure of image agreement [95% CI] across datasets for the full-mask, Ax mask, and MySh mask. Comparison between automated and manual axon counts for the training (B,C), validation (D,E) and test (F,G) data sets.