Abstract
Purpose::
To develop an algorithm to perform automated axon counting in normal optic nerves with similar or better accuracy than that obtained by trained observers (TO) and an existing semi-automated commercial software package (CSP).
Methods::
Images of myelin-stained axons from normal monkey optic nerve cross-sections are captured at 100X using a standard bright field microscope and hand-counted by 4 TOs and by a CSP (BioQuant, Nashville, TN). The same images are fed to a custom hybrid (raster and vector) algorithm (HA) that, through a series of image processing operations, separates valid axons from the background. First, a Hessian edge detection filter is applied to each original image to produce an edge map. The original image and the newly created edge map are then optimally thresholded using a Fuzzy c-means (FCM) clustering method. The two resultant binary maps are digitally combined through a binary "OR" operation and further skeletonized and pruned to obtain the outline of every closed shape (and potential axon) in the map. Once all coordinates are known, a decision tree is applied to classify each shape based on several parameters that include size, intensity, and shape complexity.
Results::
The mean of the 4 TOs axon counts for 6 normal optic nerve cross-section images were compared to the counts by the CSP and the HA. All 6 CSP and 5/6 of the HA counted images were within the 95% confidence interval. On average, the CSP counted 3% less axons whereas the HA counted 3% more axons. To further validate the HA, 362 additional normal optic nerve cross-section images were counted using the CSP and the HA. The results show an R2 of 0.9 with a slope of 1.03 and a mean difference of -17.6 ± 40.4 axons out of a mean axon count of 559.2.
Conclusions::
The results have shown that our HA is able to accurately classify axons in normal optic nerve cross-section images. By taking the regular threshold binary map one step further and describing each closed skeletonized shape as a set of (x,y) coordinates, the algorithm is able to exploit measures such as shape complexity, size, grouping characteristics, etc, to successfully classify each potential axon. The use of the FCM methods to optimally select threshold levels virtually eliminates user intervention making it possible to process large amounts of images and to accurately count 100% of an optic nerve cross-section in a short period of time.
Keywords: imaging/image analysis: non-clinical • optic nerve