Abstract
Purpose :
Grading posterior capsular opacification (PCO) varies greatly between studies and often relies on subjective interpretation. Most studies dampen the effects of intra-rater variability and grader bias by employing multiple raters; however, this increases the complexity and cost of projects, which can be avoided using software automation.
The purpose of this study was to develop software for measuring the area and intensity of PCO in post-mortem eyes, which was then compared to grading performed by trained specialists.
Methods :
Twenty-five formalin fixed cadaver eyes with intraocular lens (IOL) were obtained from the Minnesota Eye Bank. The capsular bag and IOL were removed and imaged by the Olympus DSX100 microscope then graded for PCO by two specialists using a previously described protocol. Next, our software identified the image background, peripheral, and central areas of the capsular bag. To detect PCO, we assumed the affected areas would have lower luminosity than the background. Therefore, we created a Gaussian Mixture Model of the background luminosity in the CIELab colorspace and compared the image pixels with this baseline, thus generating the PCO area and intensity for both central and peripheral areas.
To validate our results in the peripheral zone, area and intensity were directly compared to the grading performed by specialists. Validating the central zone was more complicated since the graders’ evaluation involved a single value that aggregated area and intensity. Therefore, we trained a random forest classifier to transform the generated values into a final grade. The mean square classifier error (MSE) with the expert graders was used to evaluate the classifier performance.
Results :
In the central area, the classifiers’ predictions were on average less than one grade away from the specialists’ with an MSE of 0.44 with respect to the first grader, and 0.34 to the second. In the peripheral area, the PCO area moderately correlated with both graders (R scores of 0.49 and 0.47 respectively) while the intensity strongly correlated with them (R scores of 0.64 and 0.74).
Conclusions :
Our automated PCO detection method correlated well with grading performed by trained specialists. Therefore, our system produced meaningful results and can easily be scaled for larger projects. Our next step will be to compare the detected PCO levels with a more objective and clinically relevant score, such as visual acuity.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.