Biomedical Engineering Reference
In-Depth Information
were taken using a Nidek stereo fundus camera. The original images were taken
with color slide film and then digitized with a Polaroid slide digitizer. The images
used for registration had 256-gray levels and comprise about 5% of the total
surface of the retina centered on the optic nerve head.
Following is a brief description of the image files. All image files are labeled
based on the patient identification, l/r, l/r, and a number. The first “l” or “r”
indicates that those are the images of the patient's left or right eye. The second
“l” or “r” designates the left or right stereo images. These images were taken
simultaneously onto a single photographic slide and were separated into two
images manually after digitization. The last digits, 5, 6, or 7, indicate whether
the images were taken in 1995, 1996, or 1997.
Image set 1
. There are four images in this patient file: ll5, ll7, lr5, and lr7. They
are the same images as in Fig. 4.2 [5]. These images of left eyes were taken in
1995 and 1997, 18 months apart, during which time the patient had an operation
to relieve pressure related to glaucoma. They are displayed in Fig. 4.2 here. This
image set gives us two stereo image pairs (ll5/lr5, ll7/lr7) and two temporal pairs
(ll5/ll7, lr5/lr7).
Image set 2 (B853)
. There are six images in this set: ll6, ll7, lr6, lr7, rl7 and rr7,
which gives us three stereo pairs (ll6/lr6, ll7/lr7, and rl7/rr7) and two temporal
pairs (ll6/ll7 and lr6/lr7).
Image set 3 (H3397)
. There are four images in this set: ll6, ll7, lr6, and lr7,
which gives us two stereo pairs (ll6/lr6 and ll7/lr7) and two temporal pairs (ll6/ll7
and lr6/lr7).
Image set 4 (P374)
. There are eight images in this set: ll5, ll6, lr5, lr6, rl5, rl6,
rr5, and rr6, which gives us four stereo pairs (ll5/lr5, ll6/lr6, rl5/rr5, and rl6/rr6)
and four temporal pairs (ll5/ll6, lr5/lr6, rl5/rl6, and rr5/rr6).
All together we have 11 stereo image pairs and 10 temporal image pairs,
which is a subset of the images used by Ritter
et al
. [5].
4.3.2
Mutual Information as a Measure
In this section we study mutual information as a measure for retinal image regis-
tration. Look at an extreme case first: an image registers to itself. At registration,
the pixel values in two images have an exact one-to-one (identical) relation.
In their joint histogram, there would be a straight, diagonal line. For any two
