Purchase this article with an account.
Stuti L Misra, Jonathan D Oakley, Charles NJ McGhee, Ellen F Wang, Dipika Patel, Patrick M Tarwater, Joseph L Mankowski; Automated analysis of in vivo confocal microscopy images of corneal nerves. Invest. Ophthalmol. Vis. Sci. 2017;58(8):3530. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In vivo confocal microscopy of the cornea (IVCM) allows for non-invasive acquisition of two-dimensional images at a cellular level, enabling detailed corneal nerve assessment. Manual quantification, however, is time-consuming and subjective. Automating these measures will provide clinicians with objective tools to process images promptly and provide an ‘in-clinic’ report to patients. In this study, performance of our automated approach relative to expert ground truth in normal and diseased corneas is reported.
The algorithm is an adaption for IVCM imaging of the method reported in [Dorsey et al, Am J Pathology, 2014]. Briefly, denoising is followed by edge enhancement. Processing continues on localized image patches, and in the principal orientation of the nerves. Integrating in these directions reduces the detection task to 1d while improving sensitivity to the low contrast and fragmented appearance of the nerves. Three confocal images of corneal sub-basal nerves from each of 10 patients with diabetes (M=7, F=3; age range 24-69) and 10 control participants (M=5, F=5; age range 30-73) were acquired and processed (N=60). Manual tracings from three expert readers were performed using ImageJ to obtain total nerve length and density. Bland-Altman analysis presents inter-observer reproducibility, and the Pearson coefficient gave the correlation across groups (Stata, StataCorp, College Station, TX).
Comparing the average counts across the readers and the algorithm, the Pearson coefficient reported values of 0.706, 0.862 and 0.934 for controls, diabetes and then all cases, respectively. A Bland-Altman plot (Fig 1) shows the total nerve densities readings versus the algorithm. Overall R2 correlation to individual readers was 0.89, 0.80 and 0.87 across all images.
The results show strong correlations between manual and automated counting methods. Interestingly the lowest correlation was in control cases. Further analysis showed algorithm sensitivity declined at the periphery of images, where illumination was dim. This could be addressed by cropping the image data, more uniform illumination across the images, or by algorithm improvements.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
Bland-Altman plot of nerve fiber density for the average reader score and the algorithm. R2 is 0.88, and the limits of agreement are 0.0016 and -0.0124. Diseased cases are shown as stars, controls as circles.
This PDF is available to Subscribers Only