Abstract
Purpose: :
To develop an algorithm for the automatic analysis of corneal keratocytes in confocal imaging studies.
Methods: :
Confocal microscopic images from 9 patients were assessed both manually and with a novel algorithm, which utilized morphologic, active contour, and auto-regressive techniques to identify corneal stromal keratocytes, compute their density in corneal stroma, and compute their reflectance. The manual and automatic measurements of keratocyte density and keratocyte reflectance were compared in order to assess the accuracy of the algorithm put forth.
Results: :
The algorithm achieved a sensitivity of 0.92 overall and showed no significant differences in sensitivity across four common corneal disease states. The algorithm also demonstrated a positive predictive value (PPV) of 0.84 overall and showed no significant differences in PPV across four common corneal disease states. The correlation between automatic and manual assessments of keratocyte density was found to be statistically significant (r = 0.69, 95% CI: 0.56 - 0.78, p = 5.8e-14). The correlation between automatic and manual assessments of keratocyte reflectance was also found to be statistically significant (0.74, 95% CI: 0.63 - 0.82, p = 1.11e-16).
Conclusions: :
The algorithm put forth shows sufficient accuracy and precision in keratocyte recognition across physiologic and pathologic corneal states to be of use in the analysis of confocal imaging studies of the cornea.
Keywords: cornea: stroma and keratocytes • imaging/image analysis: clinical • image processing