Purpose:
To develop and evaluate a method for intensity normalization and spatial registration of series of fundus photos for the detection of pathologies in longitudinal screening for diabetic retinopathy.
Methods:
The fundus data sets were taken from an ongoing diabetic retinopathy screening program at the Eye Hospital of Rotterdam from 2008 to 2011. An annual screening records four fields per eye.First, the intensity of the fundus images were normalized using an improved version of Foracchia's method (Foracchia et al., Medical Image Analysis 2005;9(3):179-190) that was capable of handling the bright rim at the border of some of the photos (image a and b).Second, image registration was applied to the combined set of photos from both the baseline and the follow-up screenings. For each time point, an image-mosaic was made for each eye. The multi-resolution registration algorithm consisted of a global 12 parameter quadratic model in which the order increases from coarse-to-fine. The result was then refined by local block matching (image c).Third, a difference image-mosaic was computed, highlighting the changes that occurred between the screenings (image d).
Results:
We tested the accuracy and the robustness of the method by adding noise to ten intra- and ten inter-examination image pairs. The registration error was smaller than 1 pixel (7 microns) in around 90% of the overlap. In the border region of the mosaics, the errors were slightly larger, up to approximately 3 pixels (21 microns).When adding noise, the mean absolute displacement was below the minimum size of a micro aneurysm (20 micrometers) for 92% of the registration results (image e). Only for very small overlap and large noise levels, the registration result was not good enough for further automated change detection.
Conclusions:
The robustness evaluation and subsequent visual inspection of the difference image-mosaics show that our registration method permits the detection of micro-aneurysms in longitudinal screening. After further quantitative analysis of the accuracy of our registration method, we will develop a computer aided detection system for use in longitudinal screening, aimed at automatically detecting all relevant pathologies based on difference image-mosaics
Keywords: diabetic retinopathy • imaging/image analysis: clinical • retina