Cellular in vivo 3D imaging of the cornea by confocal laser scanning microscopy

: We present an in vivo confocal laser scanning microscopy based method for large 3D reconstruction of the cornea on a cellular level with cropped volume sizes up to 266 x 286 x 396 µm 3 . The microscope objective used is equipped with a piezo actuator for automated, fast and precise closed-loop focal plane control. Furthermore, we present a novel concave surface contact cap, which significantly reduces eye movements by up to 87%, hence increasing the overlapping image area of the whole stack. This increases the cuboid volume of the generated 3D reconstruction significantly. The possibility to generate oblique sections using isotropic volume stacks opens the window to slit lamp microscopy on a cellular level.


Introduction
Presently, a well-established method for acquiring corneal images at cellular level is the combination of the Heidelberg Retina Tomograph (HRT) and the Rostock Cornea Module (RCM; both Heidelberg Engineering GmbH, Heidelberg, Germany), which is a confocal laser scanning microscope. Since its first presentation in 2002 [1], the HRT + RCM serves as a reliable instrument [2] and plays an important role in ex vivo and in vivo studies of human or animal corneas for a qualitative and quantitative analysis based on 2D imaging and/or 3D image reconstruction, e.g. the anatomical comparison of laboratory animal corneas [3], the assessment of stromal modifications of patients with progressive keratoconus after treatment by riboflavin-UVA-induced cross-linking of corneal collagen [4], quantitative full-thickness corneal 3D imaging [5], automated quantification of morphologic features of different epithelial cell layers [6], 2D reconstruction of the subbasal nerve plexus (SNP) from volume scans in the presence of ridge-like tissue deformations [7] and large-scale image reconstruction of the SNP [8][9][10][11][12][13].
Besides confocal laser scanning microscopy based methods, the full-field optical coherence tomography (FF-OCT) [14] is a promising attempt to obtain en face corneal images. Recently, FF-OCT was demonstrated to offer in vivo human corneal images with 1.7 µm lateral resolution and 1.26 mm x 1.26 mm field of view in a non-contact measurement [15]. Nevertheless, confocal laser scanning microscopy still offers higher lateral resolution and better image quality.
Several corneal 3D reconstructions from confocal light and confocal laser scanning microscopy image stacks are published already. Image stacks through the entire cornea are presented in [16] and were acquired with a tandem scanning confocal light microscope and without image alignment, leading to lower image quality and lower resolution compared to confocal laser scanning based methods. In [7,17,18], volume imaging with a confocal laser scanning microscope (HRT + RCM) was described. The internal focus drive of the HRT was used, which allows the recording of corneal stacks with a maximum depth of only 60 µm. The full thickness corneal stacks presented in [5] were performed on a sedated rabbit to avoid eye movements. The RCM's manual drive for focal plane control was replaced with a computercontrolled motor drive. One of the biggest flaws of this method and the RCM in general is the axial movement of the contact cap (TomoCap) for focal plane control, which inevitably changes the contact pressure on the cornea. From our experience, two major issues arise therefrom. First, the movement of the TomoCap away from the cornea during throughfocusing can ultimately lead to a loss of contact. Second, when the TomoCap moves towards the cornea, the increasing applanation pressure on the cornea induces compression artifacts that manifest as ridge-like deformations in the SNP and the adjacent tissue [19][20][21][22]. These deformations are particularly detrimental for imaging thin layers such as the SNP, which cannot be kept in focus over the entire field of view once the deformation height exceeds the depth of field of the RCM.
In this study, we demonstrate our in-house developed piezo-driven cornea module (RCM 2.0), which is an improved version of the original RCM. (Please note: This RCM version is not commercially available and Heidelberg Engineering was not involved in its development. The authors do not hold any intellectual property on the device or design and are currently not planning to make it commercially available.) The integrated piezo actuator is used to move the objective lens for an axial focal plane shift of up to 500 µm without moving the TomoCap. This enables precise closed-loop focal plane control. We applied our new method to demonstrate in vivo through-focusing measurements and 3D reconstruction of a human cornea.
Furthermore, we present a new TomoCap with a concave surface to reduce involuntary eye movements (so-called saccades). The functional principle is based on the fact that the eyeball's center of rotation differs from the cornea's center of curvature and on the usage of a viscous contact gel. Reduced lateral eye movements result in larger fields of view of the aligned image stack. The achieved improvements, arising from the new cap design, were assessed by analyzing comparative measurements using the standard planar and the new concave TomoCap design. Additionally, we compared three image registration methods (none, rigid, non-rigid) for the alignment of the image stacks. Orthogonal as well as oblique slices through a volume data set recorded with the concave TomoCap are presented to exemplify the capabilities of the modified RCM.

Heidelberg Retina Tomograph and Rostock Cornea Module 2.0
The HRT is a confocal laser scanning microscope primarily designed to acquire in vivo retinal images. It offers an image resolution of 384 pixels x 384 pixels with 8

Data acq
To evaluate th several corne settings) for e sets is visual stacks were ac One imag acquired usin axial distance voxel size of were scaled a the RCM is re in the adjace epithelium an Because o Therefore in a range excludi Third-part image stacks one, we used image registra which we term uses rigid tran the KIT-align rigid transform field of view approximately negligible wit defined as rig location for ea 3. True to scale sec m thickness. The c on with cr. Vertica otation of the eye sponding to a late omoCap and schem of a concave c ip objective (7 y, the diameter fore, we suppo quisition, imag he RCM 2.0 an a stacks (30 fr each TomoCap ized to demon cquired with a e stack intende ng the concave e of 1 µm betw 0.91 x 0.91 x accordingly to g eported to be ab ent images. T nd ends 187 µm of the differing a post-processi ng overexpose ty software, Im and for the im both a rigid im ation algorithm m in the follow nsformations (t nment is tailore mations. It is in horizontal l y 65 µs [7]. A th regard to ca id units. s, the size a multiple way, n st is To verify the different performance of the planar and the concave TomoCap, the error probability p for each parameter and subject was estimated using Welch's t-test. We also investigate the capability of the concave TomoCap regarding MA for stacks with large n im .

Results
As mentioned in the method section, the magnification changes with objective lens position. Therefore, an object micrometer was imaged for various objective lens positions. The total field of view was ranging linearly from about 349 µm to 351 µm in the total closed-loop travel range of 500 µm. Hence this effect can usually be neglected.
Using the standard TomoCap, involuntary eye movements can cause the cornea to move more than the width of the field of view within the exposure time of a single frame [10]. In this case consecutive images do not overlap and therefore cannot be aligned. In general, larger eye movements lead to smaller overlapping areas, thus impeding image alignment. Qualitative visual inspection of the image streams reveals a significant reduction of lateral eye movements in the data sets recorded with the concave TomoCap. To demonstrate this issue, we included six supplementary videos showing aligned (KIT-alignment) image stacks in posterior scan direction of the three subjects with both TomoCaps (Visualization 1, Visualization 2, Visualization 3, Visualization 4, Visualization 5 and Visualization 6). Table 1 shows the determined parameters MS p and MS c including the standard error for the three subjects. The mean approximated overlapping area MA p and MA c are listed in Table 2 for n st = 40 and the specified n im . Furthermore, Table 3 gives MA c for increasing n im . For comparison, the area of a single image is 0.1219 mm 2 .  Fig. 6(A). Ax in axial direc contact gel h opposite direc required for th the viscosity o It is also p axial resolutio and in [16] t direction. The orthogonal sli the epithelial determined to nt (Fig. 5(B)) s a worse align t a translation ed distortions. corrects the m g. 5(C).
the KIT-alignm n arbitrary dire xial scanning ar ction. To move as to be press ction the surro hese processes of the contact g  The imag demonstrated is difficult or appearances o ured as well, if iameters betwe possible to reli ally to Fig. 6(A ells to the st g imaging dept rficial cells (F corneal epithe cytoplasm of oplasm with b ation. The sup igure 6(C) dem ppear more un m. However, t tightly packed ght and clear c more uniform i g. 6(E)) is clea detect typical Y r. We could a cesses. An acel tuated between 5 µm. Anterior This layer reve agen fibers in-b tly higher clo a (Fig. 6(H), 6( the alignment of the differen Fig. 7).  Fig. 6(D)) ne. These m between l nucleus. brane. We g parallel with fine r limiting guished at Bowman's uclei. The seems to g towards ollowing, s will be g , .
re clearly Hence, it how clear 7(B) still exhibits some compensated Fig. 7(C). Th alignment. Al (Visualization In Fig. 8, a 8(A)) and thre is shown. Th alignment (Fi stack, it was decreased from x 292 µm (0 Table 3 Fig. 7 alization 9) em ematic drawing reconstructed c onstructed by u ( Fig. 8(D)) ag ve a cuboid v mm 2 ) to 273 µm ment and KIT-a nds on the stack ayers are prese nge highlighte s superficial ce rmediate wing containing the n Fig. 8 With prop image stack, artifacts that a 9(C)). corneal image ack is 266 x 28 to 500 µm are acquisition tim ed by a half sc r of 13 s using t te oblique slic in Fig. 9(A), th nt ( Fig. 9(B)) an with the concave T (C  Fig. 5 and Fig. 7 prove that the best image quality is realized with the KIT-alignment, which is tailored to the HRT. Compared to the planar TomoCap, the image data acquired with the concave TomoCap have the same high image quality throughout all our measurements while simultaneously reducing eye movements significantly. Although expanded in vivo human corneal image stacks and arbitrary slices were presented before [16][17][18], the image quality of the slices is now increased by the interplay of the piezo drive, the concave TomoCap and the KITalgorithm.
With the wide range of scanning parameters like velocity and travel distance as well as various scanning functions, the RCM 2.0 is suitable for numerous tasks. Furthermore, the maximum focus shift is increased to 500 µm and the position control is more precise while at the same time offering higher resolution and acceleration compared to the motorized RCM described in [5]. This enables a fast focal plane change with more than 800 µm/s while offering a position feedback with 12 nm accuracy. Additionally, the piezo system in combination with the fixed TomoCap provides new opportunities, for example a fast oscillating focal plane for large-scale image reconstruction of the SNP [29]. This is an elegant way to get rid of gaps and foreign tissue in large area imaging of the SNP [29,30], caused by unwanted axial shifts of this layer during examination. Without the fast oscillating focal plane shift, the focus has to be adjusted manually during image recording.
Despite the advantages of the concave TomoCap presented in this work, the usage of this cap is only recommended for investigations where the eye is supposed to remain fixed. For other tasks, e.g. mapping structures by the use of guided eye movements [13], the planar TomoCap is still the best choice. It is also not recommended to investigate the limbal region with such a concave TomoCap. Another drawback of the new TomoCap is its size. Often a camera is used to observe the reflection of the laser spot on the cornea for prealignment before contacting the cornea. This is only possible for large distances between cornea and the concave TomoCap, because otherwise the line of sight is blocked. However, a part of the scanning laser light is visible on the iris. This light can be exploited for a rough alignment. A multimodal imaging method combining confocal laser scanning microscopy and optical coherence tomography can also help in the alignment process. Sometimes the size of the TomoCap increases the setup effort. While examining subjects with a drooping eyelid, care has to be taken, that the eyelid is not between cornea and TomoCap.
Generally, the presented concave TomoCap and the RCM 2.0 enable fast and reliable recordings of expanded in vivo human corneal image stacks. Because of the reduced eye movements the cuboid volume of the 3D reconstruction is increased. Furthermore, whether a planar or concave TomoCap is used, the RCM 2.0 is preferable to the original RCM because of the computer controlled focus drive with a piezo actuator and the fixed TomoCap explained and demonstrated throughout this paper. The capability of the proposed method to create high-quality orthogonal or oblique slices through the corneal tissue presents the opportunity for in vivo slit lamp microscopy on a cellular level.