An improved algorithm for the assessment of quantitative vitreous haze from fundus photographs

Abstract

Purpose : To improve the quantification of vitreous haze (VH), a key clinical feature of uveitis, we modified the Passaglia quantitative vitreous algorithm (PQVHA). The PQVHA was developed to improve upon the Miami Scale, which utilizes color fundus photographs (CFP) graded by an expert reader rating the VH as an integer from 0 (no haze) to 8 (extreme haze). The PQVHA addresses important limitations of the Miami Scale: dependence on an expert reader, subjective grading and a categorical scoring system. Our modifications to the PQVHA aim to reduce its sensitivity to variability of image acquisition and quality.

Methods : The data was from a phase 2 randomized, double-masked and placebo-controlled study to evaluate the efficacy and safety of Sarilumab in 58 patients with Non-infectious, Intermediate, Posterior or Pan-Uveitis. Each of the 5 PQVHA steps were modified. 1) The green channel of the CFP had the most contrast so was used for analysis, rather than a grayscale conversion. The images were then histogram normalized and gamma corrected. 2) A Gaussian kernel was used as a high-pass filter instead of a rect function, as the latter provides non-monotonic sinc-weighting to spatial frequencies. 3) Local entropy was not used because it can provide extreme weighting to small artifacts. 4) The magnitude spectrum was used instead of the power spectrum, as the latter’s square function provides excessive weighting to a small number of high spatial frequencies. 5) The magnitude spectrum was integrated, and log transformed for a final value. Given the lack of ground truth to assess haze scores, the robustness of PQVHA and our algorithm were assessed by comparing the correlation of the scores with two expert readings in N=3245 images.

Results : The R² between our algorithm values and raters 1 and 2 was 0.33 and 0.50, respectively. For comparison, the R² between the values from the PQVHA and raters 1 and 2 were 0.10 and 0.08, respectively, and the R² between rater 1 and rater 2 was 0.34.

Conclusions : Our method builds upon the PQVHA to increase robustness against the particular experimental conditions between subjects and time points common to clinical trials.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

A rank ordering of images across all images by haze score from the PQVHA, with PQVHA haze in the top left. The haze score does not correspond to the blur the image determined by inspection.

 

A rank ordering of all images by haziness value from our modified algorithm.