Accuracy Assessment of Fundus Image Registration for Diabetic Retinopathy Screening

Abstract

 

Purpose

 

To evaluate the accuracy of a newly developed coarse-to-fine hierarchical fundus image registration algorithm for diabetic retinopathy screening and to compare it with the state-of-the-art.

 

Methods

 

Data from 70 diabetic patients who visited the Rotterdam Eye Hospital in 2012 and 2013 as part of an ongoing diabetic retinopathy screening program was included. During each visit, four images consisting of macula-centered, optic nerve-centered, superior, and temporal regions of the retinal surface were acquired from each eye. Pre-processing: the individual red-free fundus images were normalized for illumination and contrast variation between images, improving the visibility of fine vasculatures. Registration: a coarse-to-fine hierarchical registration method with an increasing deformation model complexity was applied to pairs of normalized images to create a fundus mosaic (Fig 1). Evaluation: the accuracy was assessed on the fundus mosaics resulting from our approach and the best performing method in a recent comparative study (Chen et al., IOVS, 2011), which is i2kRetina. The possible grades were: - Off: an image is placed at an incorrect location. - Not Acceptable: a misalignment worse than the width of the misaligned vessel. - Acceptable: a misalignment less than the width of the misaligned vessel. - Perfect: no noticeable misalignment.

 

Results

 

560 fundus images were combined into 140 mosaics with both methods. Two experienced graders independently assessed the accuracy of the normalized mosaic images. Each eye was graded by one grader and the order of presentation was random. The evaluation results are summarized in Table 1. The results from both graders showed a statistically significant performance difference between the two methods (Wilcoxon signed-rank test, p=0.004 and p=0.02 for grader 1 and 2, respectively). Overall, our method produces a higher success rates (83% vs. 69%,p = 0.001) of creating acceptable or perfect mosaics.

 

Conclusions

 

The accuracy assessment showed that the proposed coarse-to-fine hierarchical registration method has a significantly higher accuracy compared to the state-of-the-art. Normalization of fundus images for illumination and contrast shows much more details than the red-free images.

 

 

Fig 1. An example of (a) color fundus images, (b) red-free, and (c) normalized mosaics.

 

 

Table 1. Evaluation results of (a) grader 1 and (b) grader 2 on 105 and 35 eyes, respectively.

 

Keywords: 499 diabetic retinopathy • 550 imaging/image analysis: clinical • 549 image processing