Equalizing Uneven Illumination in a Retinal Image Using Blood Vessels as a Reference

Adam Hoover

Last revised October, 2000

Problem

The illumination in a retinal image is uneven. During the capture of the image, the amount of light falling on different parts of the retina varies depending on the direction of the illuminating flash. This direction is not known, and varies from image to image. The following eight images show examples of the variability in illumination.

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

Note: Although the images are acquired in full color, the green band of the RGB data contains the most useful information, and will be the only band examined here.

The uneven illumination prevents absolute interpretation of the intensities in the image. For instance, the optic nerve is usually the brightest object in a retinal image. Applying a simple high threshold to the image should produce pixels inside the optic nerve. However, the uneven illumination often causes the area directly under the flash to appear brightest. The following eight images are the thresholded results of the above images, showing the approximately 1,500 brightest pixels.

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

Methods

We investigate a method for equalizing a retinal image to restore the true relative illumination. The method is based upon analysis of the blood vessels. It is assumed that the blood vessels should have constant intensity across the image. A local average intensity of the blood vessels is computed for each pixel, as well as the global average intensity of all blood vessel pixels. Every pixel in the retinal image is adjusted by an amount that will make its local blood vessel intensity equal to the global blood vessel intensity.

The following images show example blood vessel segmentations, where a white pixel is designated as blood vessel. The segmentations were produced using the method reported in [1].

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

Results

The following images show the average local blood vessel intensity for the example images.

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

The following images show the results of applying the intensity adjustments to the original retinal image.

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

Conclusions

The method appears to work well. The darker and brighter areas in the original images appear more equal in illumination. The following images show the results of thresholding the brightest 1,500 pixels after this equalization has been applied.

image 0001 image 0002 image 0003 image 0004 image 0005 image 0006 image 0007 image 0008

It is clear that the optic nerve is now more easily detected as the absolute brightest object in the image. After equalization, the optic nerve is found in seven of the eight thresholded images, whereas before equalization it was only found in three of the eight thresholded images.

In some cases, certain lesions appear brighter than the optic nerve. Therefore the method of simple thresholding, even after illumination equalization, is not sufficient to detect the optic nerve. This is the case in image 0001. We are currently investigating the combination of the method presented here with that of fuzzy convergence [2], which achieved an optic nerve detection rate of 65% without searching for bright areas. We hypothesize that this combination will be successful in locating the optic nerve in greater than 90% of retinal images.

[1] "Locating Blood Vessels in Retinal Images by Piece-wise Threhsold Probing of a Matched Filter Response", IEEE Transactions on Medical Imaging , 2000.

[2] "Fuzzy Convergence", in IEEE Conference on Computer Vision and Pattern Recognition , 1998.

for more information: ahoover@clemson.edu