For each image frame, the WFS estimate of the 66 Zernike coefficients (ANSI format [19
] - 10th
order) was used to construct the corresponding wavefront aberration profile. Monochromatic light and constant amplitude across the unobstructed 6.8mm pupil diameter was assumed. By Fourier-transforming the retinal image and multiplying by the ideal optical transfer function (OTF) divided by the OTF computed from the residual Zernike coefficients the residual error and diffraction blur was removed. Where the OTF was much smaller than the ideal value, noise amplification can become an issue. To avoid this, the correction filter was clipped to a maximum complex amplitude while retaining its phase. The resulting filter was applied in the spatial frequency domain and Fourier-transformed back to the angular image space.
The resulting images were further enhanced to facilitate analysis. The lower angular-frequency albedo structure was estimated using a Gaussian high-pass filter, adjusted so as to pass the cone and rod spatial frequencies while suppressing the larger variations. Using this filter the deconvolved image was decomposed into two parts: (i) a low-angular frequency background image and (ii) a high-angular frequency image containing the cones and rods. Examination of the power spectral densities (PSDs) of the averaged images suggested that a power-law enhancement of angular frequency power would bring the rods' visibility in-line with the cones. The higher-spatial frequencies (beyond the rod PSD) had a lower signal-to-noise ratio and the power-law enhancement would boost this noise source. This was avoided by including a 2-D Chebyshev filter adjusted to pass the desired structures, while strongly suppressing the noise-dominated spatial frequencies. Once this was done, and the power normalized to preserve the cones' visibility, the resulting enhanced image was added back to the background, resulting in an enhanced retinal image. All image processing operations are linear in intensity, allowing us to use a correlation registration algorithm [9
] which corrected for small shifts between images and removed any torsional eye motion. Approximately 5-7 images were registered and summed.
shows the effect of image processing for subject N1. The same structures can be seen in the single, registered and deconvolved images.
Fig 1 Single, registered sum and deconvolved retinal images for subject N1. 650 nm, 10° TR, field of view (FOV) is 28 × 28 μm. (a) dark subtracted, single image frame (b) deconvolved and filtered image of (a). Image (c) shows the background (more ...)
shows enhanced images for subject N2 at 5° and 10° TR at 650nm. The cones show a decrease in spatial frequency with increasing eccentricity having peaks at 30 and 24 c/deg at 5° and 10° respectively. The rods showed PSD peak at 95 c/deg for each location.
Fig 2 Enhanced retinal images for subject N2. 650 nm, FOV is 28 × 84 μm. Registered sum of 5 images. (a) 5° TR, c-c cone and rod spacing were 9.3 ± 1.7 and 3.1± 0.6 μm respectively (b) 10° TR, cone and (more ...)
The mean values (15 rod or cone c-c measurements in each case) for all 4 subjects at each location are given in along with histological comparisons [10
]. The cone and rod spacings were determined by manual measurement from the 1° AO images. Additionally, PSDs were determined to confirm the cone and rod c-c spacing using the full 1° images.
Table 2 Comparison of measured photoreceptor center to center (c-c) spacing and histological data [10,11].
] is a concern for light sources with high spatial and temporal coherence. The expected speckle size for a 6.8 mm pupil is 1.9 μm at 650 nm increasing to 2.35 μm at 750 nm. A flashlamp source like the one used here is temporally incoherent but to ensure that the rod structures were not due any spatial coherence, several parameters were varied. On varying the wavelength, the same structures were observed at the same retinal locations on the same subject, this was also true when the imaging procedure was repeated 7 days later. Moreover, the separation of the rods did not change with either wavelength or time. Finally, the diameter of the entrance pupil was varied (1.5 - 3 mm diameter in 0.5 mm steps), again the same structures were observed at the same locations with the same separation. As the entrance pupil was increased a decrease in the cones and rods contrast was observed. This is due to less efficient coupling of the light into the cone receptors and hence increased scatter from the underlying choroidal structures.
Rods were not observed throughout the entire image, however collectively; the results indicate that these are indeed rods. In order to image them more readily and routinely, further improvements in imaging technology are required, e.g., utilizing pupil plane obscurations centered at the peak of an individual's Stiles Crawford function to suppress the light from the brighter cones. Utilizing the difference in directional sensitivity between rod and cones and differential bleaching may also help. Other imaging approaches such as AOSLO [21
] offers improved signal to noise and lateral resolution and the potential for averaging many more frames. Since rods are often first to be damaged in retinal disease, if one can image them as readily as we can with cones, in-vivo
monitoring of rod viability will make an invaluable clinical tool from both diagnosis and therapeutic aspects.