Both images ‘a ’ and ‘b’ have been thresholded using the same value. The difference is striking. This illustrates the importance of equalising the mean luminance of all the images prior to
thresholding. Fig. 2.29 i Thresholding of image 2.27. Mean Luminance of image was 107. • Fig. 2.29 ii Thresholding of image 2.28. Mean luminance of image shifted to 128.
2.1.4.6 Final Image Processing Steps
The use of the FFT to process the lamina cribrosa images was chosen over the unsharp masking method. In addition, image deconvolution prior to FFT was found to be useful for the HRT images.
Therefore, images were grabbed by the two systems, and underwent the following image processing method:
HRT:
1. 32 images series grabbed by HRT
2. Images converted into individual TIFF files ( HRTCONV program, HRT software)
3. TIFF files loaded into Autodeblur (version 7.5 gold. Autoquant Inc, New York, USA) and 3-D blind deconvolution of all 32 images
4. TIFF image of optic disc at the level of the neuroretinal rim chosen 5. Image magnified x2 to increase size to 512 x 512 pixels (for
comparison with Zeiss)
6. FFT of said image (Scion Image, NIH shareware)
7. Application of custom-made* annular filter to remove appropriate spatial frequencies
8. Inverse FFT
9. Histogram arithmetic to shift mean intensity 10. Thresholding
11. Blood vessel mask placed over thresholded image 12. Image analysis
Zeiss:
1. Images recorded on videotape
2. 32 frames grabbed, inspected for distortive eye movements, aligned and averaged if clear (Grabber 3, Institute of Ophthalmology)
3. Image scaled to compensate for induced curvature and magnified (section 2.12)
4. Image cropped to 512 x 512 pixels (for FFT)
5. Average image loaded into Scion Image and FFT performed 6. Custom made filter* applied
7. Inverse FFT
8. Histogram arithmetic to shift mean intensity 9. Thresholding
10. Blood vessel mask over thresholded image 11. Image analysis
*The annular filter used to eliminate ‘unwanted’ spatial frequency information was determined thus; circular filters of progressively increasing size, with their origin at the centre of the transform, were examined to see what spatial frequency information was passed. From this, it was determined what size of circular pass filter contained the most low spatial frequency information, and what size of annulus was required to contain desirable high spatial frequency information- i.e lamina pore information- without inducing artefacts. The same size annulus was used for each subject (see figure 2.1).
Table 2.2: Comparison of image processing methods for HRT and Zeiss cSLO images.
HRT
Zeiss
32 image series through depth of optic nerve head digitally recorded.
Videotape recording of optic nerve head at a single image plane. Individual images converted into TIFF
file format.
32 images digitised, aligned, averaged and saved as TIFF file format.
Blind deconvolution of image series using Autodeblur (vers 7.5, Autoquant Inc.).
Averaged image scaled to compensate for image curvature.
Image at level of optic disc chosen and magnified x2 to increase image size to 512 2 pixels.
Averaged image cropped to 512^ pixels.
FFT of image using Scion Image Software (NIH software).
As HRT.
Custom filter applied to eliminate low spatial frequency information.
As HRT.
Histogram arithmetic to shift mean intensity.
As HRT.
Thresholding. As HRT.
Blood vessel mask applied over thresholded image.
As HRT.
2.1.5 Scanning Laser Ophthalmoscopes II- Laser sources and penetration
2.1.5.1 Imaging the Lamina Cribrosa- Considerations
Anatomical details of fundus structures are more easily seen using monochromatic light, because of the different absorption and reflectance
characteristics of the stratified layers of the fundus (Delori and Gragoudas 1976; Delori, Gragoudas et al. 1977; Delori and Pfiibsen 1989). When imaging the lamina cribrosa, knowledge of the reflectance and absorption characteristics of the sclera is required, as this area closely resembles the lamina area. However, there are other ocular structures that might affect the amount of light transmitted to the lamina area and the scanning detector are the ocular media, blood vessel walls and stratified layers of the fundus, all of which have been discussed in section 1.2.3.
The aim of this initial study was to see which wavelength of laser light would be best to image the lamina cribrosa.
2.1.5.2 Wavelength of laser light and imaging the lamina cribrosa
The HRT cSLO used in this project has a single laser source of wavelength 670nm, which is capable of imaging retinal structures such as the macula and the optic nerve head. The prototype Zeiss cSLO has two inbuilt lasers, the argon (Ar; wavelength 512 nm) and helium-neon (HeNe; wavelength 632.8 nm) as described in section 2.12. An additional external near infrared (NIR; wavelength 830 nm) laser has been added, for the purpose of microperimetry testing. As it has been reported that longer wavelengths have greater penetration and scatter less (Manivannan, Sharp et al. 1993; Manivannan, Kirkpatrick et al. 1994; Eisner, Burns et al. 1996), we considered that using an 830 nm laser could give better images of the lamina cribrosa due to improved penetration. We wished to determine which of the laser sources best imaged the lamina cribrosa area, and this was investigated in the following.
Materials and Methods
The project was approved by the research ethics committee of Moorfields Eye Hospital, and performed in accordance with the 1964 Declaration of Helsinki. Two subjects were consented and had one eye dilated with guttae tropicamide 0.5%. Images of the optic nerve head were taken with the Zeiss cSLO at three wavelengths using the HeNe, Ar and NIR laser sources.
Images were recorded onto videotape and then digitised with custom-made digitisation software (Halfyard, Wade et al. 1999). The software enables individual inspection of digitised frames, for increased quality control. Frames were discarded if the optic disc appeared distorted (due to saccadic eye movements) or if the subject blinked.
To reduce noise in the images, 32 good quality consecutive frames were aligned and averaged. Digitisation, alignment and averaging of video frames were
performed using purpose written software (Wade and Fitzke 1998, Wade, 1999 #619).
Scion Image software (National Institute of Health, Maryland, USA) was used to generate the FFT of the averaged image, and a custom made filter was designed to filter out the unwanted spatial frequency information from the image. The filtered FFT was then inversely transformed to recreate a new 2-dimensional image with only high spatial frequency information within it. Section 2.3 details the image processing technique. Details of pore structure were extracted from this final, filtered image.
Image quality-variance
When assessing images, it is desirable to find the best ‘ image of a series. This is often a qualitative measure of image quality. However, we wished to develop a method that would give a quantitative measure of image quality.
In statistical analyses, the variance of a set of data is a measure of the spread of data. It is given by the formula