Evaluation and improvement of the performance of an adaptive

optics scanning laser ophthalmoscope

Ana Rita Carvalhoanarcarvalho@ist.utl.pt

Instituto Superior Tecnico, Lisboa, Portugal

October 2014

Abstract

Adaptive optics in visual science is of enormous research and clinical interest. The combinationof adaptive optics with scanning laser ophthalmoscopy (AO-SLO) provides high resolution real-timeimages of the human retina, enabling an in vivo visualization of retinal cells such as photoreceptors andblood cells. The underlying goal of this thesis is the characterization and optimization of a confocalAO-SLO which at the start of the thesis could not provide images with sufficient quality. A numberof modifications and new features were implemented within this thesis that included not only theexperimental setup but also software development in order to improve adaptive optics control. Inaddition software for off-line processing of the AO-SLO images was developed and implemented. Afterimplementation of these modifications the AO-SLO system was used to image healthy volunteers.Images of the cone and rod mosaic as well as of retinal vasculature are presented.Keywords: Adaptive Optics, Scanning Laser Ophthalmoscope, Retinal Imaging.

1. Introduction

In 1987 the first confocal scanning laser ophthal-moscope (cSLO) was developed by Webb et al.[1],whose key feature is its ability to acquire in-focusimages from selected depths using a pinhole. Thisdevice is able to deliver images of the living retinawith higher contrast and better resolution [1].

The cSLO’s ability to optically section wasnonetheless limited by eye aberrations, introducedby the optics of the eye, keeping axial resolution lim-ited to over 200 mm. A solution based on adaptiveoptics (AO) technology was reported by Roorda etal.[2], where an adaptive optics element was for thefirst time successfully implemented in a cSLO.

This successful combination of AO technologywith cSLO systems allowed the numerical apertureto be maximized, increasing light collection and im-proving both lateral and axial resolution. For theAO-cSLO, it was reported a resolution of ∼2.5 mmlateral and ∼100 mm axial, improving the typicalcSLO resolution of 5 mm lateral and 300 mm axial[2].

Many ocular pathologies arrive from retina re-lated diseases, such as retinal dystrophies whichlead to retina degeneration and eventually partialor total blindness. With this deployment of adap-tive optics technology in SLOs emerged the possi-bility to non invasively examine single cells in theliving human retina, including rod and cone pho-

toreceptor cells, retinal pigment epithelial (RPE)cells and white blood cells [1, 2, 3, 4]. It is there-fore a valuable tool for the investigation, diagnosisand follow-up of this diseases.

The underlying goal of this work is the upgradeand optimization of a confocal AO-SLO, which wasin the beginning not able to provide high qualityimages of the living retina.

2. Adaptive Optics cSLO

Adaptive Optics can be described as a method thatimproves the optical signal quality in an imagingmodality by acting actively on the optical beamwavefront, after passing through a media. A stan-dard AO system is comprised of three key compo-nents: (1) a wavefront sensor, (2) wavefront cor-rector, typically a deformable mirror (DM) and (3)real-time feedback control software [5].

In the case of eye imaging modalities, by extract-ing information about the eye wavefront monochro-matic aberrations, these aberrations can be com-pensated for in order to improve image quality [5].

Traditionally, in a confocal SLO a pinhole isplaced before the detector, in an imaging planethat is conjugated to the retina plane, in orderto reject all out-of-focus backscattered light. Theoptical sectioning property of confocal imagingimproves considerably the resolution in both axialand lateral directions. With AO technology, the

1

ability to visualize with high resolution distinctlayers of the retina (e.g. nerve fibers, bloodvessels, and photoreceptors) provided by theconfocal nature of the cSLO is further enhanced.The high contrast images delivered by AO-cSLOsystems arise from two main reasons: (1) smallerilluminated spot on the retina, which stems fromthe aberrations-corrected ingoing beam and (2)less interference from out of focus light, which isrejected by the confocal aperture [6, 7].

3. Experimental Setup

3.1. Description

The lens based AO-cSLO design is shown in Fig-ure 1, which includes both the optical pathway forillumination and collection.

The main elements comprising the AO-SLOcan be split into: light source, wavefront sensing,wavefront correction, raster scanning and lightdetection.

Light Source: The light source used for bothimaging and wavefront sensing is a superlumines-cent diode (Superlum SLD-371) operating at 842.8nm with a spectral width of 50.8 nm (full widthat half maximum). Usually, light sources in AO-SLO systems are highly incoherent in order to re-duce speckle effects on the final images, which arisefrom the interference of light backscattered from thecomplex microstructure of the retina [5]. Despitebeing scanned in a raster on the retina, the lightis descanned on the return path, resulting in a sta-tionary beam. This makes it possible to use thesame light for both wavefront sensing and imaging.Since the wavefront sensor sees the backscatteredlight as if it was coming from a single spot, it ispossible to measure aberrations [8]. One advan-tage of using the same light for both purposes isthe reduction of non-common aberrations betweenthe wave aberration and imaging light that mightreduce AO-correction performance [2].

A power reaching the eye below 700 mW is re-garded as safe for the used laser wavelength region(840 nm), according to the international lasersafety standards (EN-60825-1), even in the case ofsingle point illumination (without scanning). Thebeam diameter can be measured using a WinCamwith Data Ray software, by considering a diameterwhere light intensity drops to 1/e2 (≈ 86%) ofits maximum value. Before the collection optics abeam diameter of 8 mm was measured. At the eyea beam diameter of 6.9 mm was measured.

Wavefront sensing: The Shack-Hartmann(SH)-sensor used for this adaptive optics setup isfabricated by Optocraft and it is supplied with

the software SHSWorks. The software, which canbe driven in a Matlab or a Labview (NationalInstruments, Austin, TX) interface, allows the userto measure the wavefront and to compute differentwavefront specifications, such as the RMS. Notethat in this setup the wavefront sensing is polar-ization sensitive, in order to avoid interferencesfrom reflexes within the system, making use ofpolarization optics.

Wavefront correction: The deformable mirrorconsists of a membrane that can be deformed by97 actuators (Alpao HI-SPEED DM97-15), placedat a plane conjugate to the entrance pupil of theeye. Aberrations are corrected on both the ingoingand outgoing light paths, which contributes intwo ways for the optical quality in the confocalAO-SLO. (1) By correcting aberrations on theingoing path, light is focused to a more compactspot on the retina; (2) By correcting aberrationson the outgoing path, backscattered light fromthe retina is refocused to a compact spot at theconfocal aperture [7]. The DM was driven in a Lab-View (National Instruments, Austin, TX) interface.

Raster scanning: To get a two dimensionalimage of the retina a resonant scanner (GSI-Lumonics) with a natural frequency of ∼4 kHz isresponsible for scanning the rows (x-direction) andwhenever a row is finished a galvo scanner (model6230H from Cambridge Technology) is responsibleto adjust its angle to the next line (y-direction).Since each line of the sample is scanned in both di-rections by the resonant scanner, the line scanningrate is ∼8 kHz. The resonant scanner signal worksas the master clock for a Field ProgrammableGate Array (FPGA) unit. This unit providesnot only the driving signal for the Galvo scannerbut also the start trigger for the data acquisition [9].

Light detection: When light enters the eye itis transmitted through birefringent materials, suchas the cornea, which modify its polarization char-acteristics. Therefore, when it is backreflected fromthe retina (retinal photoreceptor layer is known topreserve light polarization) and crosses the QWPits polarization state is not exactly in a linearorthogonal state to the incident light polarizationstate. As such, and in order to maximize lightcollection at the APD, this set-up’s light detectionis polarization insensitive.

3.2. Upgrades

3.2.1 Avalanche Photo Diode

The selection of a detector in an AO-cSLO is crit-ical, since the amount of light that can be used to

2

Figure 1: Scheme of the AO-cSLO experimental setup. LS: Laser source, EC: Entrance collimator, P:Polarizer, PBS: Polarizing beam splitter, f1 – f8: achromatic lenses, A: Aperture, GS: Galvo scanner, RS:Resonant scanner, DM: Deformable mirror, DiM: Dichroic mirror, P1-P2: Prisms, QWP: Quarter waveplate, E: Eye, SHS: Shack-Hartmann sensor, APD: Avalanche photo diode.

expose the retina is limited to ∼700 mW. Adding tothis, only a small amount of light, about 1/10000 ofthe total power reaching the eye, will be reflectedfrom the retina. This reflected light is posteriorlysampled in a fine scale (less than 1 mm per pixel),which results in a very low number of photons perpixel. If we consider the quantum efficiency1 of thedetector, the photon number will be even smaller.

In this thesis two different APD modules weretested: the Si-APD-Modul LCSA500-10SMA fromLaser Components [11] and the C10508 model fromHamamatsu [10].

In order to compare the quality of the images pro-vided by each of the above mentioned APD-modulesthe signal-to-noise ratio (SNR) was calculated usinga model-eye. The model-eye consists of an achro-matic lens(f=30 mm, d=2.54 mm) and a sheet ofpaper, which represents a scattering sample. Toprevent inaccurate conclusions resulting from a po-tential saturation of the detectors, the measure-ments were performed with a lower power at theeye, around ≈ 87.7 mW.

The snoise was measured by blocking the lightpath so that no light was sent to the detector, yield-ing a value of ≈ 39 a.u. for the Hamamatsu APDand ≈153 a.u. for the Laser Components APD.

The Hamamatsu APD proved to perform better,with a SNR of 16 dB against the 13 dB calculatedfor the Laser Components APD. Therefore in thefollowing the Hamamatsu APD was used.

1Quantum efficiency is defined as the ratio of the numberof electron-holes pairs generated in the photo detector to thenumber of incident photons. It is expressed as a percentage(%) [10].

3.2.2 Shack-Hartmann Wavefront Sensor

In previous work a low sensitivity of the SH-sensorCCD camera has been reported. It was concludedthat the amount of collected electrons, resultingfrom the light returning from the retina, was be-low quantization noise2 [9]. This prevented in vivoimaging.

To overcome this limitation several independentsteps were taken. First of all, the maximum data in-tegration time (DIT)3 was increased by overwritingthe value imposed by the Matlab ALPAO Core En-gine (AO-loop software provided by the deformablemirror manufacturer) script using the PFRemote2.34 software (software provided by the CCD cam-era manufacturer). With a DIT value higher thanthe previously imposed maximum of 24000 ms it waspossible to ensure a resultant increase in the num-ber of collected photons.

The former light power reaching the eye wasaround 200 mW, which when reflected back, dueto losses within the system, reached the SH-sensorwith not enough intensity for an efficient wavefrontmeasurement. The solution was to couple morelight from the light source, which provides an out-put power of about 20 mW. The input signal iscoupled into a 90:10 coupler and the use of the90% fiber-port, rather than the 10% fiber-port, witha polarization controller afterwards allowed an in-

2Quantization noise is the quantization error producedby an imperfect transformation of analogue signals to digitalsignals [12].

3The integration time of a CCD array is defined as theamount of time that charge is allowed to accumulate [12].

3

crease up to 400 mW at the eye, while keeping itwell below the 700 mW value regarded as safe. Notethat the system is operated simultaneously with aline scanning laser ophthalmoscope which thereforelimits the allowed maximum power of the AO-cSLO.

In addition, a pellicle that was used in the oldsetup configuration to split imaging light from thewavefront sensing light of the SH-sensor was re-placed by a 50/50 beam splitter (PBS). As a result,50% of the light power is now being directed to theSH-sensor. This further improves the SNR of theSH-sensor.

A realignment of the whole AO-cSLO systemwas performed in order to eliminate a shadowthat was visible at the top of the SH-sensorduring model-eye measurements. This shadoworiginated from a partly beam blocking at one ofthe scanners. Finally, to avoid noise caused bystray light, a cover was built for the SH-sensor.All together these procedures improved signifi-cantly the quality of the wave-front sensing andcorresponding wavefront reconstruction, which inturn yielded not only a more accurate correction ofthe eye aberrations but also a more stable AO-loop.

4. System Software

4.1. Description

Both the deformable mirror and SH-sensor wereprovided by ALPAO which also supplies the AL-PAO Core Engine (ACE) software, that can be fur-ther customized to the particular needs of the cus-tomers. The ACE is implemented in object orientedMatlab code enabling the user to derive classes fromsupplied classes and overwrite different class meth-ods [13]. Briefly, this software allows monitoring ofthe SH-sensor’s CCD camera raw image, monitor-ing of the reconstructed wavefront and root meansquare (RMS) value of the residual wavefront errorand control of the DM.

However, the software’s base modules can not bechanged and are not designed for ophthalmic appli-cations. Due to this it was not possible before thisthesis to record images of the living retina with theAO-cSLO setup with satisfactory quality. Withinthis thesis several software adaptations in the AO-loop related software have been performed to enableretinal imaging. In addition new software featureswere implemented to improve the control of the de-formable mirror.

When initializing both the DM and the SH-sensor, a circular mask (used for spot detection inthe SH-sensor) is defined as well as the wavefrontreconstruction algorithm (modal or zonal). Aftersetting this definitions the AO-loop control is per-formed using a Labview interface, where the usercan see in real time the raw image from the SH-

sensor as well as the wavefront reconstruction andrespective RMS value. Apart from that two newfeatures were added. Now the Labview interfaceenables recording of the reference raw image of theSH-sensor to be subtracted in Matlab, whose pur-pose will be further explained later. In addition,the second implemented feature allows recording ofthe raw SH-sensor image, the corresponding wave-front reconstruction and the RMS wavefront error.This enables monitoring of the AO-loop efficiencyand residual wavefront error for both the model eyeand in vivo measurements.

4.2. Upgrades

4.2.1 AO-Loop Software

One major problem of the original adaptive opticsloop software was an instability of the loop in thecase of in vivo measurements, which turned out tobe mostly due to an inefficient computation of thespots’ center of gravity (COG) in the SH-sensor.

The center of gravity computation of each spotof the SH-sensor is a key process for the adaptiveoptics loop. To eliminate noise from the spot detec-tion a threshold level is predefined. Pixels that showvalues below this threshold are set to zero and areexcluded from the centroid calculation. In addition,centroids are only calculated within a circular maskthat corresponds to the maximum pupil size. Un-fortunately, this threshold is implement as a fixedparameter in the software and can not be changed.

While there is sufficient signal in the case of themodel-eye to detect every centroid within the pupilmask, only few were detected in the case of in vivomeasurements (cf. Fig. 2(a)). Due to the lowsignal-to-noise ratio, many of the detected COGsare erroneous. The software automatically extrap-olates the wavefront based on the found centroidsover the full mask and the AO-loop tries to correctfor this erroneous wavefront. This leads to arbitraryshapes of the DM, a completely blurred AO-SLOimage and finally to a saturation of the elements ofthe DM, stopping the AO-loop since the elementsare frozen and can not be moved.

In order to overcome this problem, the class forthe SH-sensor was replaced. The new overwrittenclass allows for a subtraction of a reference imageto the raw image of the SH-sensor, in order to elim-inate any external undesirable influences such as re-flections within the system or stray light that can bedetected by the sensor. In this new class it is nowpossible to set a gain. Since the threshold couldnot be changed, this gain allowed an amplificationof the signal such that the majority of the signalpixels lay above the threshold, allowing a correctcentroid calculation. The enhancement in the COGcomputation after this software change is depictedin figure 2(b).

4

(a) (b)

Figure 2: Spots computation in the SH-sensor dur-ing an in vivo measurement in two different situa-tions. (a) illustrates the few spots being calculatedbefore applying the gain correction and (b) showsthe enhancement in the number of spots being cal-culated after the software changes.

Originally, the reconstruction of the wavefrontwas based on the modal reconstruction algorithm(polynomial fitting using a set of Zernike polyno-mials), which extrapolates values in regions of thefull mask even in the case where the sensor doesnot find centroids. For instance, when performingin vivo measurements the eye pupil varies betweenindividuals, which means that in some cases (wherethe pupil is smaller) not all lenslets within the maskare illuminated. Hence, the wavefront is extrapo-lated to provide values within the full mask. Thisextrapolation is sensitive to errors and causes wrongdisplacement of the corresponding elements of theDM. This leads to an incorrect wavefront correction,image blurring and finally to an unstable AO-loopwhich will result in saturated DM elements and astopping of the AO-loop. To avoid this error onepossible approach is changing the reconstruction al-gorithm from ”modal” to ”zonal”.

(a) (b)

Figure 3: Wavefront measurements in the model-eye in a case where the pupil of the model-eye issmaller than the predefined mask. Wavefront re-construction based on the modal (a) and the zonal(b) algorithm. Severe differences in the wavefrontcan be observed between zonal and modal recon-struction.

The zonal reconstruction is based on an iterativeprocess, whose tolerable error can be changed. Fig-

ure 3 illustrates the difference between zonal andmodal reconstruction for a pupil that is smaller thanthe mask.

4.2.2 Deformable Mirror Software

Two additional features were implemented in thedeformable mirror software.

Bias Vector: In the beginning of any measure-ment the voltage value being applied to each ac-tuator of the DM is zero, therefore the shape of themirror surface is completely arbitrary. A flat mir-ror shape would be the ideal initial condition priorto the start of the AO-loop since this random shapewill cause additional computational time and strokeof the DM to correct the real eye aberrations. Inorder to have the mirror in a flat position wheninitializing the AO-loop a predefined actuator com-mand vector is set as ”biasVector”. To record thepattern of the DM that corresponds approximatelyto a flat surface a highly reflective surface was po-sitioned at the location of the eye and the AO-loopwas turned on. The corresponding voltage valuesof each actuator were saved and are now automati-cally applied to the DM every time the loop startsto run.

Additional Defocus: In the healthy eye the maincontribution to the SH-sensor signal comes from thephotoreceptor layer. Therefore, this layer will beimaged sharply when the AO-loop is turned on. Inorder to image other layers (e.g. anterior layers suchas the nerve fibre layer) an additional defocus hasto be introduced to the beam, which can now beeasily done using the a LabView interface whichwas adapted for this purpose.

4.3. Image Post Processing

Custom software was developed in Labview (Na-tional Instruments, Austin, Texas) for the image ac-quisition and for the several offline processing steps,illustrated in figure 4.

The AO-cSLO records an image frame consistingof 2569 × 1096 pixel in ≈ 69 ms, yielding a framerate of ≈ 14.6 frames per second (FPS).

The amplitude and number of lines scanned bythe Galvo-scanner can be set in the program Scan-nerControllerfor2.vi whereas the amplitude of theresonant scanner is set by adjusting a potentiome-ter.

The resonant scanner moves the beam over thesample and image data is recorded during bothscanning directions (forward and backward). As aconsequence of the sinusoidal waveform of the res-onant scanner, the images need to be dewarped.However, in order to obtain a full image, every sec-ond row (backward scan) has to be inverted and aresidual displacement has to be corrected. Figure

5

(a) (b) (c)

Figure 4: Standard image processing steps for AO-cSLO images. (a) Raw image data of the forward andbackward scan of the resonant scanner. (b) Single frame after dewarping and interwoven. (c) AveragedImage after applying correction to the signle frames. The AO-cSLO image size is 2569 pixel×1096 pixeland covers a field of view of approximately 330 mm×330 mm. Scale bars = 20 mm. Image location is ∼ 8°

nasal from the fovea.

4 shows the two images resulting from the forwardand backward scan and the interwoven image.

In a next step the recorded images have to beregistered to each other in order to allow for imageaveraging.

When imaging living human eyes, normal invol-untary eye motion, even during fixation, causesthe imaging raster to move continually across theretina. Eye motion consists of drift, tremor and sac-cades. Saccades are very fast and result in heavilydistorted image frames which are afterwards elim-inated. Although slower, drift and tremor are fastenough to cause not only displacements between theframes but also in frame. Therefore the image reg-istration software includes two steps.

First of all, a reference frame that does not showmotion artifacts is selected from the recorded im-ages. In a first step displacements between the ref-erence frame and the other recorded frames is deter-mined using cross correlation between them. Thisinformation is used to correct for the lateral dis-placement. Since intra-frame distortion exist in allimage frames, each frame is in a second step di-vided into subframes (stripes) along the slow scan-ning axis (each consisting of 40 horizontal lines).Cross correlation was then calculated between eachstripe and the reference frame and the correspond-ing residual displacement of a stripe was determinedand corrected. With this, such intra-frame distor-tion within the data set was eliminated. After thatthe cross correlation between the frames was cal-culated again and only frames that showed a highcorrelation with the reference frame were used foraveraging. Averaging an undistorted sequence offrames is an important step to improve the signalto noise ratio. The number of frames contributingto an averaged image varied from 10 to more than30. The final averaged image is shown in figure 4(c).

5. System Characterization

5.1. Field of view

To evaluate the field of view (FOV) of the AO-cSLOan U.S Air Force 1951 resolution positive test tar-get (RTT) from Melles Griot was used. It consistsof a series of several bar patterns, which have a sim-ilar shape but different gradually smaller sizes. Amodel-eye, consisting of a lens with a focal lengthof 30 mm and the RTT representing the retina, wasbuilt and images were recorded using the AO-SLOsystem (cf. Fig.5).

Figure 5: Resolution test target imaged with theAO-cSLO.

These images were processed and different pat-terns with known physical extension were selected.The size of each pattern was determined within theAO-cSLO image by counting the number of pixels inorder to determine the average pixel size. A FOV of558 mm×570 mm was calculated, which correspondsto a scanning angle of ∼ 1◦ × 1◦.

5.2. Resolution

Different approaches were used to evaluate the sys-tem resolution. A calculation of the lateral resolu-tion provided by the AO-cSLO was first performed,after which an analysis based on acquired imageswas made.

6

5.2.1 Diffraction limit

Assuming a well designed diffraction-limited oph-thalmoscope, the lateral resolution is only limitedby the width of the diffraction-limited point spreadfunction (PSF), which can be calculated as follows:

d =1.22 × 0.840 × f

n × 6900(1)

In the above equation the constant variables arethe beam diameter at the eye (6.9 mm) and thebeam wavelength (840 nm).

Considering the optics of the eye, and assuminga refractive index (n) of 1.33 and a focal length (f)of the eye lens of 22.22 mm, a lateral resolution of2.5 mm can be calculated [14].

For the model-eye the same calculation can beperformed assuming f=30 mm and n=1, whichyields a value of 4.46 mm for the lateral resolution.

In this calculation the improvement in the reso-lution resulting from the confocal configuration ofthe AO-cSLO is not taken into account. Instead, alarge pinhole (larger than one airy disk diameter) isassumed.

5.2.2 Resolution test target

The resolution limit of the AO-cSLO was evaluatedusing the same model-eye, consisting of a 30 mmlens and the resolution test target.

Figure 6: Zoomed region of interest of the resolutiontest target imaged with single-mode fiber configu-ration. The red lines mark the lines whose intensityprofile was analysed in order to confirm the systemresolution capabilities.

The smallest element of the test target in the AO-cSLO image that can be seen defines the lateral res-olution, that is typically expressed as the reciprocalof the spatial frequency of this element pattern. Byobserving the resulting image (cf. Fig. 5), one canconclude that the smallest resolved element is No.4 of group 7 for both x and y directions. This wasfurther investigated by analysing the intensity pro-files for the pixels along the red lines, marked as redlines in Figure 6, which confirmed that the system

was able to resolve this element in both directionsby showing three well defined peaks.

For this group the manufacturer of the RTT pro-vides values of 181 line pairs per millimetre whichcorresponds to a lateral resolution of ∼ 2.76 mm inboth directions.

Assuming that the ratio between the resolutionwith the eye and with a model-eye is approximatelyconstant, one can estimate the resolution capabili-ties of the system when imaging the eye using themeasured value with the RTT. As such, the esti-mated resolution is ∼1.55 mm. Note that this res-olution is better than the diffraction limit withoutusing a pinhole.

6. Results and Discussion

6.1. Imaging of foveal cones

Cone photoreceptors act like tiny waveguides.Therefore, light that enters them is preferentiallyredirected back through the pupil, producing a highcontrast mosaic of bright spots in the retinal image.Therefore they have been one of the favourite tar-gets in AO-cSLO imaging [15].

In figure 7(a) the individual foveal cone photore-ceptors can be clearly resolved almost throughoutthe whole image, with exception of the central partof the fovea (lower left corner, marked with anarrow). Because of the large variation of the re-flectance of each photoreceptor it has been previ-ously proposed to display these images in a logarith-mic scale [16]. The dark lines visible in the imagesare shadows of vessels, whose location is anterior tothe confocal image plane.

Due to the regular arrangement of the photore-ceptors, it is possible to quantify their spacing (i.e.next neighbour distance (NND)) by performing aFast Fourier Transform (FFT). The dominant spa-tial frequencies resulting from the photoreceptor’sregular arrangement of the photoreceptor appear asrings on the FFT, the so called Yellott rings. Theradius of the rings will then correspond to the rowto row spacing of the individual photorecptor type.The visibility of the Yellott’s rings can be regardedas a good indicator for the capability of an AO-equipped instrument to resolve individual cells suchas rods and cones. As such, the 2D FFT was calcu-lated for the selected region assigned with a whitesquare (cf. Fig. 7(a)). Yellott’s ring can be clearlyobserved in the FFT of the image (cf. Fig.8), indi-cating the regular arrangement of the cones in thisregion and confirming the high lateral resolution ofthe AO-cSLO instrument.

Assuming an hexagonal packing of the conesone can obtain the closest neighbour cone spac-ing by a straightforward multiplication of theradius value by a factor of (1/cos 30°). Thisfactor arises from geometric considerations on the

7

(a) (b) (c)

Figure 7: Reflectance images of the human cone mosaic near the central fovea from the subject1, collectedusing 840 nm light and single-mode fiber. The same image is shown with linear (a) and logarithmic (b)grayscales, to facilitate visualization of the cone mosaic (scale bars are 20 mm). The arrow points to thefovea centralis. (c) Enlarged view of the ROI marked in (a) (scale bar is 5 mm).

cone mosaic arrangement [17]. A value for theclosest neighbour cone spacing of 3.74 mm wascalculated, which is in good agreement with re-cently reported results [18] and data from histology.

Figure 8: Power spectra obtained by 2D FFT of7(c), where Yellott’s ring is clearly visible (scale baris 0.3 cycles/mm).

6.2. Imaging of rods

The size of the rods is in the order of the diffrac-tion limited resolution (∼2 mm) that is supportedby the optics of the eye. In order to visualize thesestructures perfect wavefront correction has to beensured. With the purpose of testing the abilityof the AO-cSLO instrument to resolve rod photore-ceptors the temporal region of the eye was imaged.Figure 9 shows first imaging results of the photore-ceptor mosaic that were recorded at about 7° eccen-tricity temporal from the fovea. It shows sparselypacked bright large spots (corresponding to cones)that are surrounded by smaller and densely packedspots (corresponding to rods).

In comparison with recently published high res-

Figure 9: Image of the photoreceptor mosaic show-ing rods and cones recorded in the temporal region.The orange circle marks a cone photoreceptor andthe blue circles rod photoreceptors. The scale baris 20 mm.

olution images of the rods [4] the following differ-ences between the AO-cSLO systems should be em-phasized. In [4] a wavelength of 680 nm was usedfor imaging while in this thesis the instrument wasoperated at 840 nm. Due to the longer wavelengtha 24% loss of resolution is expected which degradesthe rod visibility. In addition, no drugs for prevent-ing accommodation were applied. Therefore adap-tive optics correction in the measurements shownabove might be degraded by residual accommoda-tion influences.

6.3. Imaging of retinal vasculature

Because of its confocality, the AO-cSLO is ableto axially section different regions of the retina.The possibility to change the focal plane within theretina can be achieved by applying a certain amountof defocus on the deformable mirror.

Figure 10(a) shows an example of an image

8

(a) (b)

Figure 10: (a) Adaptive Optics SLO image with the focal plane set to the anterior layers of the retina(a) and with the focal plane set to the photoreceptor layer of the retina (b). The scale bars are 20 mm.

recorded from the retina with the focal plane ofthe AO-cSLO located at the anterior layers. Theblood vessel appears as a continuous bright line withneighbouring dark areas. The central part of thevessel is narrower, which can be explained by thefact that only the central part of the vessel scatterssufficient light back to the confocal aperture. Out-side of this central part the light is mainly scatteredinto different directions. The surrounding structureof the vessel can mainly be attributed to the retinalnerve fiber layer which shows strong backscatter-ing compared to other anterior retinal layers. Todemonstrate the ability of the system to set the fo-cal plane at different depths within the retina, figure10(b) shows images of approximately the same loca-tion but with the focal plane set to the photorecep-tor layer. The photoreceptor mosaic and shadowscaused by light absorption within the vessels can becan be clearly observed.

7. Conclusions and Future Work

Within this thesis, an adaptive optics scanning laserophthalmoscope (AO-SLO) was optimized and itsperformance was evaluated. One key feature of thesystem is the simultaneous recording of line scan-ning laser ophthalmoscopy (LSLO) images whichgreatly simplifies subject alignment and the deter-mination of the location that is imaged with theAO-SLO. After implementation of several improve-ments and upgrades the LSLO and AO-SLO cannow be operated simultaneously and the adaptiveoptics loop runs stable, even in the case of in vivomeasurements. The imaging capability of the sys-tem is demonstrated by in vivo imaging of the hu-man cone and rod mosaic as well as the visualiza-tion of retinal capillaries. The lateral resolution ofthe device was measured to be ∼ 2.8 mm which isin the line of previous reported values of resolution

for AO-SLO devices in the literature [2]. The mea-sured field-of-view of ∼ 1◦ × 1◦ is also in the rangeof values reported for AO assisted ophthalmologicdevices which are typically very small.

As future developments, the adaptive optics looptime has to be decreased. It has been reported thata closed-loop bandwidth of around 1 Hz-2 Hz shouldbe enough to correct for the eye aberrations in or-der to achieve diffraction limited imaging over di-lated eyes [8]. However, currently 3 seconds areneeded for one iteration step, which is not accept-able for patient measurements. One reason for thisrelatively slow computational time is that the usedframe grabber for the SH-sensor can only be oper-ated in a 32 bit computer environment. Modern andfast computers are using a 64 bit operating systemwhich would accelerate the process. However, thisrequires hardware changes such as replacing the SH-sensor. In addition, the exposure time currently re-quired for the SH-sensor may be reduced. Currentlythe SH-sensor samples the wavefront by a factor of∼20 more accurate than it is possible to change thewavefront (which is determined by the number of el-ements of the deformable mirror). By reducing thenumber of lenslets more photons will contribute toeach centroid which enables to reduce the exposuretime of the SH-sensor and an improvement of theAO loop speed.

Another limiting factor of the current system isthe rather slow AO-SLO imaging speed. The imag-ing speed is mainly determined by the frequencyof the used resonant scanner which is operated at4kHz. Other research groups use faster resonantscanner (e.g. 8 kHz or 16 kHz) which would im-prove the imaging speed by more than a factor of2.

Regarding the post-processing software there isalso room for improvement. Recent studies ex-

9

plored alternative techniques for motion correction[19, 20] that might be translated to the AO-SLOimages acquired with this instrument. The tech-niques are based on B-spline elastic image registra-tion, which might be easily adapted through the useof public domain software such as ImageJ.

Finally, it would be interesting to perform mea-surements in patients with known diseases, in or-der to attest for the device robustness and validateits clinical potential. Patient measurements requirethe implementation of a fixation target and the in-clusion of an additional anterior segment monitor.This monitor will be essential to facilitate the align-ment of the patient’s eye in respect to the AO-SLOinstrument.

These are just few steps that would certainly en-rich this work and take it to a next level.

Acknowledgements

This work was supported by the Austrian ScienceFund (FWF project P22329-N20).

References

[1] Robert H Webb, George W Hughes, and Fran-cois C Delori. Confocal scanning laser oph-thalmoscope. Applied Optics, 26(8):1492–1499,1987.

[2] Austin Roorda, Fernando Romero-Borja,William Donnelly III, Hope Queener, ThomasHebert, Melanie Campbell, et al. Adaptive op-tics scanning laser ophthalmoscopy. Optics Ex-press, 10(9):405–412, 2002.

[3] David R Williams. Imaging single cells in theliving retina. Vision research, 51(13):1379–1396, 2011.

[4] Alfredo Dubra, Yusufu Sulai, Jennifer L Nor-ris, Robert F Cooper, Adam M Dubis, David RWilliams, and Joseph Carroll. Noninvasiveimaging of the human rod photoreceptor mo-saic using a confocal adaptive optics scanningophthalmoscope. Biomedical optics express,2(7):1864–1876, 2011.

[5] JASON Porter, H Queener, J Lin, KARENThorn, and ABDUL Awwal. Adaptive opticsfor vision science, 2006.

[6] Francesco LaRocca, Al-Hafeez Dhalla,Michael P Kelly, Sina Farsiu, and Joseph AIzatt. Optimization of confocal scanning laserophthalmoscope design. Journal of biomedicaloptics, 18(7):076015–076015, 2013.

[7] Austin Roorda. Applications of adaptive opticsscanning laser ophthalmoscopy. Optometry andvision science: official publication of the Amer-ican Academy of Optometry, 87(4):260, 2010.

[8] Heidi Hofer, Pablo Artal, Ben Singer,Juan Luis Aragon, and David R Williams. Dy-

namics of the eyes wave aberration. JOSA A,18(3):497–506, 2001.

[9] Gerold Christian Aschinger. Adaptive opticsscanning laser ophthalmoscope. Master’s the-sis, Techische Universitt Wien, 2012.

[10] Hamamatsu Photonics. APD Module C10508,2010.

[11] Laser Components. APD ModuleLCSA/LCIA-Series, 2012.

[12] Patrick Martinez. A practical guide to CCDastronomy, volume 8. Cambridge UniversityPress, 1998.

[13] Alpao. ALPAO CORE ENGINE rev 1.3, 2012.

[14] Orin Packer and David R Williams. Light, theretinal image, and photoreceptors. The Scienceof Color, 2nd ed SK. Shevall, ed. Optical Soci-ety of America, Elsevier (Oxford, UK), pages41–102, 2003.

[15] Pooja Godara, Adam M Dubis, Austin Ro-orda, Jacque L Duncan, and Joseph Car-roll. Adaptive optics retinal imaging: emerg-ing clinical applications. Optometry and visionscience: official publication of the AmericanAcademy of Optometry, 87(12):930, 2010.

[16] Alfredo Dubra and Yusufu Sulai. Reflec-tive afocal broadband adaptive optics scanningophthalmoscope. Biomedical optics express,2(6):1757–1768, 2011.

[17] M Pircher, RJ Zawadzki, JW Evans,JS Werner, and CK Hitzenberger. Simul-taneous imaging of human cone mosaic withadaptive optics enhanced scanning laserophthalmoscopy and high-speed transversalscanning optical coherence tomography. Opticsletters, 33(1):22–24, 2008.

[18] Franz Felberer, Julia-Sophie Kroisamer, Bern-hard Baumann, Stefan Zotter, Ursula Schmidt-Erfurth, Christoph K Hitzenberger, andMichael Pircher. Adaptive optics slo/oct for3d imaging of human photoreceptors in vivo.Biomedical optics express, 5(2):439–456, 2014.

[19] Akihito Uji, Sotaro Ooto, Masanori Hangai,Shigeta Arichika, and Nagahisa Yoshimura.Image quality improvement in adaptive opticsscanning laser ophthalmoscopy assisted capil-lary visualization using b-spline-based elasticimage registration. PloS one, 8(11):e80106,2013.

[20] Johnny Tam, Joy A Martin, and Austin Ro-orda. Noninvasive visualization and analysis ofparafoveal capillaries in humans. Investigativeophthalmology & visual science, 51(3):1691–1698, 2010.

10