Abstract
Purpose:
To develop a quantitative method for grading the “blurriness “ of ocular fundus images in order to permit automated grading of the severity of vitritis.
Methods:
Several different image processing algorithms were written on the Matlab platform to quantify blurriness, which included Fourier spatial-frequency analysis, wavelet transforms, and entropy filtering methods. The algorithms were refined and validated using a set of 8 reference images that were acquired by masking a single “standard” fundus picture with a series of analog spatial filters. After reference set validation, the algorithms were applied without modification to a dataset of clinical images. The dataset consisted of 12 TIFF digital fundus images of 12 eyes of 12 patients with uveitis. Computer-scored results were then compared in a masked fashion with the subjective readings of an expert clinician (BM).
Results:
Spatial frequency, wavelet transform, and entropy filtering algorithms by themselves all performed well for the reference set, giving scores that scaled in proportion to the known blur factor, but much less so for the clinical set owing to the large inherent variation among patients and image takers. Better overall performance was achieved with an algorithm that combined spatial frequency and entropy filtering. The Pearson correlation coefficient between the mixed-method scores and physician grades of blurriness was 0.81 (R2 = 0.66) for the clinical dataset.
Conclusions:
We have developed a computer algorithm for grading vitreous haze in an unbiased and quantitative manner that correlates strongly with subjective readings of an expert clinician. We aim to now expand these findings to larger sets of fundus images and larger pools of experts.
Keywords: 550 imaging/image analysis: clinical •
763 vitreous •
557 inflammation