An autonomous classifier for nuclear cataract

Purpose: Development of a computer-based autonomous system for grading of nuclear opacification and color. Methods: The classifier was developed using digitized 35mm slides acquired with a Topcon SL5-E slitlamp camera. One set of 96 slides (2 each of 48 eyes) was chosen for minimal brunescence and to span the range of 0-4 on the Wilmer grading system for nuclear severity. A second set of 48 slides was chosen to span a range of colors from white through dark yellow (-1 to +1 on the Wilmer color grading scheme). Our procedure consisted of: (1) specification of the region of interest, (2) image pre-processing, (3) extraction of metrics, and (4) use of these metrics to yield opacification and color grades. Regression-based and neural classifiers were used. Performance was compared with that of the Wilmer grading scheme; test/retest and inter-observer effects were explored. Results: For nuclear opacification, performance against the Wilmer grading system was very good (R=0.94, n=96). For subjective test/retest of this set, we found R=0.97, n=48 with 95% confidence intervals of ±0.53 of an integer grade. The autonomous classifier yielded R=0.99, n=48, with 95% confidence intervals of ±0.28. For evaluation of nuclear color, the classifier performed well against the subjective grading. For a color grade of -1 we found a blue/red color ratio of 0.96±0.08, n=20; for color grade +1, we obtained a ratio of 0.70±0.06, n=13. Conclusions: We have developed and tested a sensitive, reproducible classification system for nuclear cataract. Such a system will prove useful in a clinical environment to aid in the diagnosis of cataract severity, and in the research environment for evaluation of drug efficacy and the etiology of cataract.