First steps toward 3D high resolution imaging using adaptive optics and full-field optical coherence tomography

First steps toward 3D high resolution imaging using adaptive optics and full-field optical coherence tomography

Leonardo Blancoa, Marie Blavierab, Marie Glancad, Florence Pouplardb, Sarah Tickb,

Ivan Maksimovicab, Guillaume Chénegroscd, Laurent Mugniercd, François Lacombeae, Gérard Roussetad,, Michel Pâquesb, Jean-François Le Gargassonb, José-Alain Sahelb

aLESIA, Observatoire de Paris, CNRS, UPMC, Université Paris Diderot, 5 Place Jules Janssen, 92190 Meudon, France

bCentre d’Investigations Cliniques du CHNO des Quinze-Vingts, Université Paris VI, 28 rue de Charenton, 75571 Paris cedex 12, France

cDOTA ONERA, BP 72, 29 avenue de la Division Leclerc, 92322 Châtillon cedex, France dGroupement d’Intérêt Scientifique PHASE (Partenariat Haute résolution Angulaire Sol Espace)

between ONERA, Observatoire de Paris, CNRS and Université Paris Diderot eMauna Kea Technologies, 9 rue d'Enghien, 75010 Paris, France

ABSTRACT

We describe here two parts of our future 3D fundus camera coupling Adaptive Optics and full-field Optical Coherence Tomography. The first part is an Adaptive Optics flood imager installed at the Quinze-Vingts Hospital, regularly used on healthy and pathological eyes. A posteriori image reconstruction is performed, increasing the final image quality and field of view. The instrument lateral resolution is better than 2 microns. The second part is a full-field Optical Coherence Tomograph, which has demonstrated capability of performing a simple kind of “4 phases” image reconstruction of non biological samples and ex situ retinas. Final aim is to couple both parts in order to achieve 3D high resolution mapping of in vivo retinas.

Keywords: adaptive optics, eye retina, high resolution imaging, optical coherence tomography

1. INTRODUCTION In vivo retinal cells studies and early diagnoses are severely limited, due to the lack of resolution on eye-fundus images resulting from “classical” ophthalmologic instruments. This defect is due to ocular aberrations time-depending variations in any retinal imager entrance pupil1. This shows that their correction cannot be achieved even through individual custom glasses. Adaptive Optics (AO) is a real-time technique well-suited to correct for atmospheric turbulence impact in astronomical images formation. AO has now been used for more than a decade to compensate for ocular aberrations too, allowing the conception of new kinds of high-resolution retinal imaging instruments2,3,4.

To pursue the goal of reaching the diffraction limit on dilated eyes, our AO retinal imaging instrument (flood imaging type) previously described5 is permanently under progress. The routine examination of healthy eyes has allowed us to identify some of its limitations. The instrument and post treatment algorithms will be described in the next session. Results and limitations will then be reported.

A second instrument is currently being developed to improve axial resolution, representing the next step to a 3D high resolution imager. Optical Coherence Tomography (OCT) is an interferometric technique allowing the non-invasive extraction of sectional images of the sample under study. There have been different approaches to OCT, some of them aiming at suppressing the scan in one or two directions, principally to save time. Thus, full-field OCT produces en-face oriented images, the only remaining scan direction being the axial one. Its theoretical axial resolution is limited by the illumination source bandwidth. Description of the full-field OCT part with its inherent difficulties will be done in the 3rd section. Early results on a non-biological sample will also be displayed.

2. AO INSTRUMENT REPORT 2.1 Setup description

Our AO system is made of a 32 x 32 subpupil Shack-Hartmann wavefront sensor (WFS) matched to a 52-actuator magnetic deformable mirror (MIRAO). The AO system operates at a loop frequency of 7 Hz. Images are taken with a 12 bit 1360 x 1036 pixel Q-Imaging Retiga camera. The imaging frequency is 7 Hz too. Retinal illumination is based on flashes (exposure duration: 1 ms to 10 ms) achieved through a system of 2 shutters synchronized with the servo. Analysis channel and imaging one are synchronized, so that the imaging flash is taken right after the mirror voltages computation and actuator stabilization. Total delay between WFS acquisition and image acquisition is roughly 85 ms (WFS Acquisition = 30ms, WFS CCD reading = 30ms, voltages computation ~5ms, mirror actuator and membrane stabilization ~20ms). The wavelength chosen for the analysis is 835 nm (Superluminescent Diode), whereas imaging wavelength is 550 nm (mercury vapor arc lamp). The choice of a reference source in the near IR allows the subject to feel quite comfortable during the aberrations reduction. The choice of a short wavelength for the imaging leads to a higher theoretical resolution than the selection of a higher wavelength. Besides, blood vessels visualization capability is better in the green part of the spectrum than in the red one. The whole AO and imaging system is controlled through a LabVIEW interface.

Fig. 1. Simplified scheme of the AO Imaging System

The fixation target is made of a 7 leds in-line stick, distant in angle from 0.43° (on the retina) from each other. It can rotate around the first led. Every led can be switched on alternatively, so that a total field of ca. 6 degrees in diameter centered on the foveola can be covered (the individual image FOV being 1°). Calibration is performed to link the ON-led to retinal image localization on left and right eyes. Looking for instance at the subject’s commercial OCT sections before AO examination allows one to select the positions of the fixation led in order to image retinal zones of interest.

Focalization on the plane of interest (i.e. cones, vessels…) is performed by translating the imaging camera along its axis. We perform coarse pre-focalization by asking the subject to look at the imaging CCD camera through a green filter. We translate the camera along its axis. Once the subject sees the CCD sharply, we assume, since his accommodation has been blocked, that the photoreceptors layer and the imaging CCD are optically conjugated. After that step, fine focalization on the photoreceptors layer is done, looking at the images (or at the images spectrum), also by translating the imaging camera along its axis.

In order to optimize image illumination density on ametropic subjects, the retinal illumination can be “tuned” in focus. By changing the distance between 2 optical elements, a more or less converging beam can be sent toward the subject’s retina. The aim is to properly illuminate the layer of interest, in order to get the highest signal-to-noise ratio on the images. Subject’s visualization of the fixation target and of the reference source for the AO can be optimized in the same way.

Wavefront and mirror data from the AO system (wavefront slopes, Zernike coefficients, WFS subpupil intensities, actuator voltages) are stored so that we can analyze those data versus the quality of the images obtained. We are currently working on this analysis.

2.2 Automatic reconstruction algorithm based on the fixating target

During an imaging sequence (typically 5 or 10 flashes), the subject is asked to fixate the led that is ON. The selection of this led in the line determines the image eccentricity on the retina. Between sequences, another led can be switched on, in order to relax the subject’s fixation. One AO imaging examination is constituted of a varying number of imaging sequences (from 10 to 30 or 40). Individual images recorded by our imaging camera are 12 bit 500 x 500 pixels images. Since the subject eye is continuously moving, the imaged zone on the retina is not exactly the same from one image to another. This is a drawback in that we cannot easily sum the individual images to improve signal-to-noise ratio (SNR). We first have to register the images before doing so. The registration algorithm is composed of four steps.

As the position of the image on the CCD may have changed from one imaging sequence to another, we first have to register the field of the images (i.e. the circular illuminated area of the CCD). We compute the centroid of each image in the stack and we shift the images so that their centroid is roughly at the same position.

The second step is an image selection one. We analyze the power spectrum of each image m ( ) (circular mean of the FFT of a 256 x 256 portion of the images) in the stack of images recorded by our camera for a given eccentricity and a given plane of interest. Images having too weak a power spectrum at the spatial frequencies corresponding to the expected cones spatial frequencies are discarded. The criterion C used is given by:

(1)

where f1 ~ 0.15 µm-1 and f2 ~ 0.35 µm-1, i.e. the expected cones frequency band at the given eccentricity.

Images with , where is the highest criterion in the stack of images and an adjustable threshold (so we can select more (to further improve SNR but with images of less quality) or less (only really sharp ones) images), are discarded. This is a relative criterion. We compare the images to the “best” image in the stack of images as opposed to an absolute reference.

Fig. 2. Top: Normalized Power Spectrum of a “good” image (solid line) and a “bad” image (dotted line). Abscissa units are µm-1

(expected cones spatial frequencies 0.15 – 0.35 µm-1). Bottom left: bad image. Bottom right: good image.

Once we have selected the good images from the stack, an analysis based on systematic cross-correlations calculations of the images (we have to carefully compensate for the non-uniform illumination of our images so that the registration method actually registers the photoreceptors as opposed to any illumination pattern or artifact) allows one to roughly derive the respective position of one retinal image with regard to the others. Images with too weak a correlation with others are automatically rejected (i.e. there are two image selection steps). The obtained accuracy is ½ pixel, i.e. 0.3 micrometers on the retina. Once the images are roughly registered we perform a second, fine, sub-pixel registration with a maximum likelihood estimator. This method was first developed to accurately register faint galaxies images6 and we adapted it to our samples. Once the sub-pixel shift between images has been computed, the images are shifted and summed (taking the variable number of individual measurements per pixel into account). This registration method is implemented in IDL.

A comparison between simple cross-correlation registration and sub-pixel registration is shown on figure 3. As the difference is not clearly visible on the registered images, we computed the improvement in the power spectrum of the same 512 x 512 pixels area on both images, which is a good indicator of the registration accuracy.

Fig. 3. Top left: pixel registered image. Top right: sub-pixel registered image. Bottom: improvement in power spectrum (in

percentage) from pixel registration to sub-pixel registration. Abscissa units are µm-1.

Improvement in the power spectrum is significant at the spatial frequencies corresponding to the retinal cones (i.e. 0.15 – 0.35 µm-1). The main drawback is that the sub-pixel registration is slower than the simple pixel registration (typically twice slower for a 10 image registration).

As our biological images are rigid and already show a good SNR thanks to the use of AO, two reconstruction options are possible. The first one turns eye’s movements into an advantage in order to get a field wider than the one intercepted in individual frames. From series of registered images, one obtains a reflectance average image, taking the variable number of individual measurements per pixel into account. Signal-to-noise ratio is increased as square root of the number of

stacked images. Thus, the field of view of the instrument is increased, up to about 4 degrees (instead of 1°). This option should prepare the way to morphological studies. The second option consists in keeping only the common part in a series of successive images. The blood flow temporal fluctuations can so be studied in each zone of the field that has several times been illuminated. The aim is to measure blood velocity in the smallest retinal capillaries, where no Doppler techniques can be applied. Using “full field” AO provides here high-resolution images without distortion, which is a first step towards functional imaging. For the moment, these reconstruction algorithms are run after an imaging sequence, but their inclusion in the imaging software is planed. An example of wide-field images obtained with the reconstruction algorithm is displayed on figure 4.

2.3 AO results and limitations

As an example of performances achieved with this system and to illustrate the wide-field reconstruction, Fig. 4 (left) shows one sequence of retinal images in the same eye (subject MP, left eye). The eccentricity is about 2 degrees and the size of each individual frame is 1° (~300 µm on the retina). Fig. 4 (middle) shows the images kept for the reconstruction using correlation algorithms. One can notice that more than 50% of the images are good enough to be used for the final wide-field image. On this subject, the actual quality of correction was better than 50 nm RMS during all the imaging process. This performance has been regularly reached on other healthy eyes.

Fig. 4. Reflectance images of one subject’s photoreceptors. Left: one sequence images. Middle: images kept for the

reconstruction. Right: wide-field image, obtained by using the reconstruction algorithms.

Limitations of the instrument are above all linked to the closed-loop frequency, which does not allow us to properly deal with some subject’s eyes movements.

2.4 Future developments for the AO imaging setup

Our next slight improvement will be to perform two AO loops per image (AO frequency = 15 Hz, imaging frequency = 7.5 Hz). This is already possible with our setup and the last version of the AO control software.

The next step will be to further improve AO loop frequency by changing the WFS camera and improving the AO control software. We believe that a faster loop (typically 10 times faster i.e. ~70 Hz) will allow our AO bench to obtain photoreceptors images on a broader population.

3. FULL-FIELD OCT INSTRUMENT The current axial resolution of our AO setup is about 40 µm. In order to improve the axial resolution of tomographic sections and to increase the instrument sensitivity, the final objective is to couple adaptive optics with optical coherence tomography (OCT) in order to build a 3D high resolution imager. Before combining both techniques, a new instrument dedicated to eye imaging is being developed, based on time-domain full-field OCT (FFOCT). With this specific OCT bench, limitations and best solutions for eye tomography will be studied.

3.1 Setup description

Since we obtain 2D retinal images with our AO setup, we decided to develop a system based on FFOCT. Unlike “classical” OCT, FFOCT presents the advantage to not require lateral scanning but just in-depth axial one7. Our time-domain FFOCT system is based on a Linnik interferometer with lenses focusing the light beam in each arm respectively on a reference mirror and on the sample under study (Fig. 6). XY cross-sections are reconstructed from the interferometric images between sample and reference acquired by a CCD camera. The source is a 770 nm laser diode with a 50 nm spectral bandwidth. Theoretical axial resolution in an OCT setup is given by light source spectral features. For a source with a Gaussian spectral distribution, its coherence length ΔL in air space is given by:

(2)

where λ is the central wavelength and Δλ the spectral bandwidth. The source has typically a 10 µm coherence length, which determines a small coherence depth in the sample. Interference fringes are then visible over this coherence length, with a maximal contrast for the zero optical path difference. Theoretical axial resolution is defined as the half of the coherence length.

The setup is controlled through a Labview software to synchronize the CCD camera and a motorized translation stage for depth scanning. Presently, reference is fixed while sample is moved; it will eventually be inverted.

3.2 Classical reconstruction algorithm for FFOCT

A phase-shift algorithm with four images is used to reconstruct en-face cross-sections inside of the sample, recombining 4 interferometric images delayed by π/2. Each interferometric image is represented by:

with m=0,1,2,3 (3)

where Iref is the intensity reflected by the reference mirror, Icoh is the coherent intensity retro-reflected by the coherence volume within the sample, Iincoh is the whole incoherent intensity backscattered by the sample, δ is the optical path difference and λ the central wavelength.

A classical phase-shift algorithm gives tomographic cross-sections by calculating:

(4)

Reconstructed images are directly proportional to the signal of interest Icoh, representing sample reflectivity at the depth defined by the zero optical path difference. Thus, full-field OCT performs directly optical slices, whose width is linked to the source coherence length.

3.3 Innovative 4 phases configuration

Because of fast eye's movements, the required 4 interferometric images cannot be recorded temporally one after the other by modulating the optical path difference, as it is realized in classical FFOCT. Our setup uses linearly polarization to acquire the needed images simultaneously on 4 different areas of the CCD camera8. Detection channel includes a Wollaston prism and a quarter-wave plate to register at the same time four π/2-shifted interferograms spatially separated. Indeed, the light beam is separated in two by a non-polarizing beamsplitter cube. One of the resulting beams travels through a quarter-wave plate and is then π/2 phase-shifted. Both beams are horizontally separated and go through the Wollaston prism. This latter separates each beam in two vertical components shifted by π (Fig. 5). With this system, there is no need for any path difference modulation in order to obtain the four π/2 phase-shifted interferometric images. This phase-shift system is mounted after the interferometer (Fig. 6).

Fig. 5. Scheme of the phase-shift system in order to obtain simultaneously four π/2 phase-shifted interferometric images

(NBPC: non-polarizing beamsplitter cube; QWP: quarter-wave plate; WP: Wollaston prism; Δ: phase-shift).

This innovative phase-shift system presents specific difficulties in comparison with classical temporal path modulation. Since the interferometric images are subtracted from each other, images must be centered one another with sub-pixel resolution and with normalized intensity. Images centering is realized with the same algorithms than for astronomical AO treatment and reconstruction is performed in IDL. Moreover this method requires calibration of the CCD camera and a homogeneous flux to avoid reconstruction artifacts. Indeed, flux normalization between each interferometric image is required in order to suppress differential offset. Otherwise, fringes contrast is hidden, signal is lost and artifacts are created. Offset can be due to inhomogeneous CCD response that can be easily corrected by measuring the camera flat-field. However intensity differences also appear between the 4 images because beams do not follow exactly the same path; they go through different optics or through different places in the same component. Differences have to be minimized by a good alignment. Nevertheless, as optics are not ideal, there will be always small differences between the 4 images. A calibration procedure is thus required to ensure reconstruction validity.

Fig. 6. Scheme of the whole polarized full-field OCT setup (PBC: polarizing beamsplitter cube; NPBC: non-polarizing beamsplitter

cube; QWP: quarter-wave plate; HWP: half-wave plate; WP: Wollaston prism).

Fig. 7. Photography of the FFOCT setup.

3.4 First OCT experiments

The OCT setup has first been tested on non-biological samples to confirm the detection and contrast extraction method. Validation of the 4 phases reconstruction from interferometric images acquired simultaneously was realized with 3D topography of a high diffusive plastic stair. Its steps are 5 µm high. This sample was chosen for the very weak light level returned towards the CCD camera. In the same way, the OCT setup will have to detect weak signals from biological tissues. The stair structure has been observed, which has proven setup ability to reveal tiny structures and to reconstruct 3D samples with an axial resolution better than 5 µm. Reconstruction algorithm, including images centering and normalization, was then validated. A cross-section and a 3D reconstruction image are shown Fig. 8.

Fig. 8. Left: a topographical cross-section of the stair, calculated from 4 interferometric images. Steps inside the coherence length are visible. Right: 3D volume-rendering image of the whole stair computed with the software ImageJ from a stack of cross-sections.

3.5 Future developments for our OCT setup

Ex situ retinal images are currently being acquired. Afterwards different biological samples will be imaged under various conditions to optimize the OCT setup. The aim is to approach in vivo eye imaging conditions. Indeed, the system has to achieve a sufficient signal-to-noise ratio, be robust and fast enough regarding eye's movements. Finally AO system and OCT setup will be coupled to perform 3D high resolution mapping of in vivo retinas. 2D lateral high resolution will be provided by AO correction while OCT will perform optical slicing.

4. CONCLUSION A 3D retinal imager is being developed, based on combination of full-field AO and OCT. Two systems are currently being built by separating AO and OCT. Our AO system has reached in its current version a level of accuracy, which qualifies it for routine clinical use. Residual wavefront errors as small as 50 nm RMS have been obtained on healthy dilated subjects. A clinical study on 240 eyes is under progress at the Quinze-Vingts Hospital. A total of 40 healthy volunteers and 200 patients in a variety of retinal pathologies will be assessed. Our OCT system is currently under development. It uses a method for extracting the interferometric contrast without any need for optical path modulation. Validation of this principle has been performed on a non-biological sample. Both setups are being separately optimized for eye imaging. Coupling both methods will permit to acquire high resolution cross-sections of in vivo retina.

ACKNOWLEDGMENTS

Authors thank A. Perchant and T. Vercauteren from Mauna Kea Technologies for their essential contribution to the reconstruction algorithms. We thank Imagine Eyes for their assistance concerning the closed-loop components.

REFERENCES

[1] Castejón-Mochón, J. F., López-Gil, N., Benito, A. and Artal, P., "Ocular wave-front statistics in a normal young population," Vis. Res., 42, 1611-1617 (2002).

[2] Hofer, H., Chen, L., Yoon, G.Y., Singer, B., Yamauchi, Y. and Williams, D.R., "Improvement in retinal image quality with dynamic correction of the eye's aberrations," Optics Express, 8, 631-643 (2001).

[3] Chamot, S. R., Dainty, J. C. and Esposito, S., "Adaptive optics for ophthalmic applications using a pyramid wavefront sensor," Optics Express, 14, 518-526 (2006).

[4] Zhang, Y., Poonja, S. and Roorda, A., "MEMS-based adaptive optics scanning laser ophthalmoscopy," Optics Letters, 31, 1268-1270 (2006).

[5] Glanc, M., Gendron, E., Lacombe, F., Lafaille, D., Le Gargasson, J. F. and Léna, P., "Towards wide-field retinal imaging with adaptive optics," Opt. Comm., 230, 225-238 (2004).

[6] Gratadour, D., Mugnier, L. and Rouan, D., "Sub-pixel image registration with a maximum likelihood estimator,” Astronomy and Astrophysics, 443, 357-365 (November 2005).

[7] Dubois, A., Grieve, K., Moneron, G., Lecaque, R., Vabre, L. and Boccara, A., C., "Ultrahigh resolution full-field optical coherence tomography," Applied Optics, 43(14), 2874-2883 (May 2004).

[8] Glanc, M., Lafaille, D., Lacombe, F., Vabre, L., Lévecq, X. and Chateau, N., "Full-Field AO-assisted OCT for high- resolution tomographic imaging of the retina," Photonics West (SPIE) 2006.