Abstract
Purpose: :
To develop an automated algorithm to count axons in NHP optic nerve images with varying levels of axonal damage.
Methods: :
420 images from 10 NHP optic nerves ranging in damage from mild to severe were categorized into 4 damage levels and manually counted by a trained observer. Using a previously developed neural network, the same images were independently classified as "normal" or "damaged." Images classified as normal by the network were then counted using our existing normal algorithm (NA). The rest of the images were processed using a new damage counting algorithm (DCA) utilizing a Canny-Deriche edge detection filter. Maximum gradient magnitudes were preserved. A Freeman Chain Code was then applied to the edge detected images to extract all closed and continuous edges. Closed edges that had an average gradient magnitude and average within-pixel intensity below one standard deviation from the mean of all closed edges in the image were further removed from the final result. The number of remaining closed edges determined the number of axons present in the image. Correlations to hand-counted images were made using statistical analysis software.
Results: :
Of the 420 images, 150 were determined by the trained observer to have a mild level (M) of axon damage, 131 moderate (Mod), 109 high (H), and 30 severe (S). Based on the output from the existing neural network, 81% of (M), 36% of (Mod), 8% of (H), and 7% of (S) images were counted using our existing NA. Counts using the NA and DCA for all images correlated well (R2 = 0.87) with the counts from the trained observer. Correlation coefficients for the M, Mod, H, and S groups of images were 0.77, 0.61, 0.64 and 0.45, respectively (p<0.001 for each). The DCA consistently reported higher counts than the trained observer for images with greater axonal damage and lower counts for images with less axonal damage.
Conclusions: :
The DCA performed well in images with mild axon damage. As axon damage increases, image texture changes dramatically with artifactual ‘axon-like’ features becoming more common. These were, for the most part, eliminated by the logic employed in our new DCA, but some residual false positives remained. Further refinements will enable the algorithm to improve the false positive rejection rate.
Keywords: image processing • imaging/image analysis: non-clinical • optic nerve