Retinal Imaging Imaging the human retina has always held considerable interest in ophthalmology and vision science. In vitro histology images have revealed millions of cells in the retina. Non-invasive, in vivo imaging makes the study of healthy retinal structure and function possible. It permits early detection and diagnosis of retinal diseases such as age-related macular degeneration, diabetic retinopathy and glaucoma. Doctors use ophthalmoscopes and fundus cameras to non-invasively image the human retina in vivo. However these low magnification, wide field of view images cannot be used to detect diseases in their early stage, to detect microscopic retinal damage or to study the cellular structure of a healthy retina.

Figure 1. Adaptive optics permits nearly diffraction limited, cellular resolution imaging of the human retina in vivo [1]

Need for superresolution in retinal imaging There are several structures in the retina that are smaller than 2 microns in size such as rods, foveal cones, fine blood capillaries, ganglion cell axons and dendrites, etc., which have been observed in vitro, but are not easily resolved in vivo in most instruments. The wavelength used for imaging and the numerical aperture of the imaging system are the two factors that limit the resolution of the system. Reduction of wavelength is not possible beyond eye-safety limits. The limiting aperture in imaging the retina is the pupil of the eye. It is not possible to dilate a patient’s pupil to greater than 6 – 8 mm diameter. Hence the spatial frequencies that are captured by the optical imaging system are limited by the size of the pupil. In order to capture the higher spatial frequency structures in the retina, it is essential to somehow effectively increase the size of this aperture. The technique of structured illumination imaging has been used in microscopy [3] to image spatial frequencies beyond the cutoff of the system. We apply this approach to a flood-illuminated adaptive optics retinal imaging system to capture those higher spatial frequencies which normally do not lie within the passband of the system to obtain superresolved images. Structured Illumination for superresolution This is a method of resolving spatial frequencies of an object that are normally outside the passband of an imaging system. Usually, the object being imaged is illuminated by a flat, uniform illumination. In this technique, instead, the object is illuminated by a sinusoidally patterned illumination. Effectively, sinusoidal illumination multiplies with the spatial frequencies of the object giving sum and difference frequencies. If the difference spatial frequencies are lower than the cutoff of the imaging system, they can be captured as aliased, moiré patterns in the image. If we have prior knowledge of the sinusoidal illumination spatial frequency then the unknown object spatial frequencies can be calculated from such images. Thus, sinusoidally patterned illumination can be used to image spatial frequencies beyond the diffraction cutoff of an imaging system.

Figure 2. Concept of structured illumination for superresolution

The effective object is the product of the object and the sinusoidal illumination pattern (Figure 2). Since the Fourier transform of the sinusoidal pattern is three impulses the Fourier transform of such an incoherent image contains three replicas of the object spectrum. The two shifted object spectra contain parts of the object spectrum which normally lie outside the passband of the imaging system. Therefore, these contain the superresolution information we desire. We can recover the three object spectra as long as we take three or more images with distinct phase shifts in the sinusoidal illumination pattern (Figure 3).

Figure 3. The three superimposed object spectra can be recovered from three or more sinusoidally patterned images

The shifted versions of the recovered object spectra are moved to their actual positions in spatial frequency space and added to the conventional unshifted object spectrum with appropriate weighting to obtain a superresolved image. The spectrum of such a reconstruction is shown in Figure 4.

Figure 4. Reconstructed image spectrum with 75% superresolution in one orientation

If a high spatial frequency sinusoidal pattern is used the reconstructed image can have as much as twice diffraction limited resolution for objects with linear absorption and emission. The above simulation uses a sinusoid of spatial frequency at 75% of cutoff spatial frequency. Therefore the reconstructed image shows 75% superresolution. It will have superresolution along the direction perpendicular to the sinusoidal pattern used. Similar reconstructions with the orientation of the sinusoid rotated by 60˚ and 120˚ can be added to the above to fill in the entire OTF (Figure 5).

Figure 5. Superresolved image spectrum with 75% superresolution in all orientations versus spectrum of conventional image taken with uniform illumination

In Figure 6 we see a comparison of (a) the pristine object, with (b) the conventional image taken with a uniform illumination, (c) the reconstructed image with 75% superresolution and (d) a comparable image taken with a uniform illumination through a 75% larger pupil. This shows that this approach really does obtain true superresolution information.

Figure 6. Comparison of pristine object, conventional image taken with a uniform illumination, reconstructed image with 75% superresolution and image taken with uniform illumination using a 75% larger pupil.