Image Reconstruction of the Subbasal Nerve Plexus with In Vivo Confocal Microscopy

Abstract

Purpose.: To overcome the anterior corneal mosaic (ACM) phenomenon in in vivo confocal laser scanning microscopy (CLSM) and to reconstruct undistorted images of the subbasal nerve plexus (SNP), facilitating morphometric analysis in the presence of ACM ridges.

Methods.: CLSM was performed in five healthy volunteers. An original image processing algorithm based on phase correlation was used to analyze and reduce motion distortions in volume scan image sequences. Three-dimensional tracing of the SNP was performed to reconstruct images containing only the SNP layer, with nerve fibers clearly visible even in ACM areas.

Results.: Real-time mapping of the SNP revealed the presence of ridges with K-structures underneath them in all cases. The occurrence of K-structures correlated directly with development of ACM observed by slit lamp and resulted in massive deformation at the level of Bowman's membrane, seriously interfering with examination of SNP structures. The average elevation of ACM ridges was 20.6 μm (range, 8.7–34.0 μm). The novel method presented permitted reconstruction of the SNP layer in regions of ACM.

Conclusions.: The described method allows the precise analysis and elimination of motion artifacts in CLSM volume scans, in conjunction with the capability to reconstruct SNP structures even in the presence of severe ACM. The robustness and automation of the described algorithms require ongoing development, but this will provide a sound basis for extended studies of corneal nerve regeneration or degeneration and for use in clinical practice.

In vivo confocal laser scanning microscopy (CLSM) allows for the analysis of tissue at a cellular level in laboratory animals ^1,2 and humans ^{3 –5} in ophthalmology. Several research groups have presented results describing the lid structures, ⁶ conjunctiva, ^7,8 cornea, ^9,10 and even the lens. ⁴ In particular, the subbasal nerve plexus (SNP) is a structure that has been the subject of ongoing discussion. Alterations of the SNP have been described in keratoconus, ^11,12 after refractive surgery ^9,13 and corneal grafting, ¹³ and in patients with diabetes mellitus ^14,15 and Fabry disease. ¹⁶ The correlations between the density, tortuosity or number of nerve fiber branches in the SNP, and corneal sensation are clinically important. To date, there have been only limited systematic and reliable in vivo studies of these structures. A review by Patel et al. ⁹ analyzed the density of the SNP measured with confocal microscopy based on 13 in vivo studies: the resulting density was strongly dependent on the type of microscope and ranged from 5.8 to 8.4 mm/mm² with tandem scanning and 0.6 to 15.2 mm/mm² with slit scanning to 21.6 to 25.9 mm/mm² measured with CLSM. Erie et al. ³ attributed this discrepancy to the differences in field brightness and image contrast between instruments. 

Two principal problems present themselves when imaging the SNP. First, the small investigated area (usually 400 × 400 μm) does not provide the necessary basis for statistically firm conclusions concerning morphometric plexus alterations. To overcome this problem, we have described elsewhere an online procedure for mapping the SNP. ¹⁷ The second problem arises from ridge-like deformations at the level of the SNP. They affect the layers of basal epithelial cells, the SNP, and Bowman's membrane and are induced by the compression applied to the cornea through the applanation technique of the microscope. Reports of the origin of these ridge-like deformations can probably be dated back to Bowman, who described in 1857 an “anterior elastic lamina from which there pass off a number of fibers into the layers of the cornea proper.” ¹⁸ In 1968, Bron and Tripathi ¹⁹ introduced the term anterior corneal mosaic (ACM), which refers to both a ridge and a groove pattern that can be induced on the epithelial surface and to an identical pattern seen deeper in the epithelium through retroillumination. He suggested “that the underlying basis of the ACM is a reticular structure, having elastic properties situated within or in close proximity to Bowman's membrane.” ¹⁹ Recently, Kobayashi et al. ¹⁰ observed a network of ridges at the level of Bowman's membrane and epithelium using CLSM and attributed the development of these ridges to the presence of bundles of fibrous structures measuring 5 to 15 μm in diameter, termed Kobayashi structures (K-structures). They also suggested a correlation of the described network of ridges with the ACM phenomenon. ¹⁸ The detection of K-structures with confocal microscopy is in line with previous light/electron microscopic observations and two photon-generated second-harmonic studies that located anterior collagen fiber bundles running along the posterior surface of the Bowman layer. ²⁰ According to Kobayashi et al. ¹⁰ and Bron and Tripathi, ¹⁹ the network of ridge-like deformations at the basal epithelium and Bowman's membrane and other manifestations of ACM are induced into the corneal tissues above K-structures by compression of the cornea. Because of the intentionally limited focal depth of confocal microscopes, ridge-like structures in the SNP region make it impossible in general to keep the SNP in focus across the entire image area. In summary, such deformations prevent the evaluation of the subbasal nerve structures. 

Images devoid of Bowman's membrane deformations are essential both for visual and for software-based detection and qualitative evaluation of corneal nerves. Consequently, there exists a major incentive to develop methods that reconstruct the authentic corneal nerve pattern from images that include artifacts. We present here a novel approach to overcome the ACM phenomenon and to reconstruct undistorted images of the SNP both for qualitative and for possible further quantitative morphometric analysis. 

Materials and Methods

In Vivo Confocal Laser Scanning Microscopy

In vivo CLSM was performed (Heidelberg Retinal Tomograph II in conjunction with the Rostock Cornea Module [HRTII-RCM]; image size 400 × 400 μm, 384 × 384 pixels, 8-bit; Heidelberg Engineering, Heidelberg, Germany; equipped with a water contact objective 63×/0.95 W, 670 nm, ∞/0; Zeiss, Jena, Germany). The distance between corneal surface and objective is kept stable using single-use contact elements (TomoCap; Heidelberg Engineering) with a planar surface, as described elsewhere. ²¹  

Investigations were performed in five healthy associates (four men, aged 32 years [AZ], 33 years [SA], 44 years [OS] and 52 years [BK], respectively, and one woman, aged 26 years). The study was conducted after approval by the ethics committee of the Medical Faculty, University of Rostock. The research followed the tenets of the Declaration of Helsinki. The study was explained in detail and informed oral consent was obtained before any investigative procedures were performed. 

After topical application of anesthetic (Proparakain-POS; Ursapharm, Saarbrücken, Germany), a drop of eye gel (Vidisic; Bausch & Lomb/Dr. Mann Pharma, Berlin, Germany) was applied into the inferior conjunctival fornix to serve as a coupling medium. The contralateral eye fixed a light target to stabilize the patient's gaze. Real-time mapping of the SNP in the central cornea ¹⁷ was performed on an area up to 1600 × 1600 μm (1536 × 1536 pixels, 8-bit) in all five healthy volunteers to permit large-scale visualization of the SNP. The duration of microscopy was between 1 and 5 minutes (Fig. 1). 

Figure 1.

 

ACM phenomenon. (A) Real-time mapping of the SNP, with visible K-structures (arrows). (B) Slit lamp photography (cobalt blue filter, Na-fluorescein) showing anterior mosaic structures (arrow).

Subsequently, image acquisition in the same region was performed in z-scan automatic volume scan mode (30 images, volume depth 60 μm, constant interslice distance 2 μm). The acquisition time for a single stack was 4 seconds. CLSM was performed in the region of interest (i.e., at the level of the basal cells, SNP, Bowman's membrane and anterior stroma) at depths ranging from approximately 30 to 90 μm. Multiple scans were performed on each subject with a total duration of microscopy up to 15 minutes. 

Processing of Confocal Images

After in vivo CLSM, the image data acquired from z-scans were processed with an originally developed digital image processing algorithm to reconstruct an undistorted image of the SNP. The new method can be divided into three successive steps: image registration, volume reconstruction, and nerve fiber layer extraction. 

Image Registration.

Each two-dimensional (2D) image (400 × 400 μm, 384 × 384 pixels, 8-bit) represents an en face optical section of the cornea. Systematic analysis of the image sequences reveals the presence of distortions (Fig. 2B) that can be consistently explained by movements of the observed eye. These cannot be completely suppressed during microscopy. Even though good patient compliance can reduce voluntary eye movements to a minimum, uncontrollable involuntary eye movements always occur. The fastest of those, so-called microsaccades, can have peak velocities of up to 120° per second (approximately 24 μm/ms in the scanning plane) and amplitudes of up to 2° (approximately 400 μm or one image length in the scanning plane). ²² The contact of the single-use contact element (TomoCap; Heidelberg Engineering) with the cornea reduces, but does not eliminate, eye movement. 

Figure 2.

 

Two successive in vivo confocal microscopic images of the basal epithelial layer (volume scan mode; interslice distance, 2 μm). (A) No eye motion during image acquisition, with no discernible distortions. (B) Extreme distortion effects caused by fast eye motion are visible in the lower half of the image.

Rigid image registration approaches do not yield good results with motion-distorted image sequences. To reconstruct the acquired volume accurately, we have therefore developed an original image registration algorithm that automatically corrects the motion-induced distortions. To understand its design, it is important to know the image acquisition process of the HRTII-RCM and the source of motion-induced distortions. An image is scanned pixel by pixel and row by row from top to bottom. Crucially, the time required to obtain one intensity value is negligible (<0.2 μs, explaining why practically no motion blur is present in the images) and the time required to scan a single row is still very short (∼65 μs). Only minimal movement can occur in this amount of time. On the other hand, it takes approximately 25 ms to scan an entire image, meaning that the last image row is scanned almost 25 ms after the first one. Any motion occurring over this period will lead to a varying shift of subsequent image rows in relation to the first row. Our objective is to establish a shift vector for each image row and to invert it to reconstruct undistorted images. 

Every image I_n is divided into several subimages, or slices, S_n,k , each consisting of a constant number of pixel rows of I_n . Using a phase correlation function, shift values x_n,k and y_n,k are calculated for each subimage S_n,k in relation to the preceding image I_n-1 . Shift values for single image rows are then estimated by piecewise polynomial interpolation between the support points defined by the subimage translations x_n,k and y_n,k . Because these values are always relative to the preceding image, they have to be transformed into a global coordinate system. The results are then used to reduce motion-induced distortion from the images. Although a rigid, translation-only approach is used for subimage registration, the described registration algorithm as a whole performs nonrigid, elastic image registration. In general, it therefore does not retain straight image borders for the motion-corrected images. 

Volume Reconstruction.

To reconstruct the acquired volume, the registered, motion-corrected images are simply stacked up and shifted according to the registration results. Each of the equidistantly acquired images produces a single layer of the resulting voxel model. 

Nerve Fiber Layer Extraction.

The following step is performed to extract the nerve fiber layer: perpendicular sections of the reconstructed volume reveal that the nerve fiber layer has a significantly higher reflectivity than its immediate surroundings (i.e., Bowman's membrane and basal epithelial cells). The extraction step begins in a central y-z section image of the reconstructed volume, tracing the brightest pixels starting from a manually provided seed point inside the nerve fiber layer. Subsequently, the tracing continues along the x-axis. The result can be presented as a depth map, which enables an image of the SNP to be assembled from the voxel data. 

It should be noted that several processing parameters currently have to be adjusted manually, such as the seed point for tracing and the subimage size (i.e., the number of image lines that constitute each subimage). 

All described image processing algorithms and visualization of most images during the process (motion-corrected images, section images, and reconstructed image of the SNP) were developed in-house and implemented in C++. The Interactive 3D Surface Plot plug-in for ImageJ software (developed by Wayne Rasband, National Institutes of Health, Bethesda, MD; available at http://rsb.info.nih.gov/ij/index.html) was used to visualize the depth map in three dimensions (Figs. 3A, 3B, 4A, 4B, and 5A, 5B). 

Figure 3.

 

Depth map and nerve fiber layer extraction at SNP level. (A) Depth map of SNP layer inside the reconstructed volume. (B) Depth map of SNP layer, textured with reconstructed image. (C) Reconstructed image of the SNP. Image size: 362 × 347 pixels, 377 × 361 μm, ∼85% of original image size.

Figure 4.

 

Depth map and nerve fiber layer extraction at SNP level. (A) Depth map of SNP layer inside the reconstructed volume. (B) Depth map of SNP layer, textured with reconstructed image. (C) Reconstructed image of the SNP. Image size: 363 × 372 pixels, 378 × 388 μm, ∼92% of original image size.

Figure 5.

 

Depth map and nerve fiber layer extraction at SNP level. (A) Depth map of SNP layer inside the reconstructed volume. (B) Depth map of SNP layer, textured with reconstructed image. (C) Reconstructed image of the SNP. Image size: 347 × 393 pixels, 361 × 409 μm, ∼92% of original image size.

Results

In Vivo CLSM

The source confocal images had largely uniform luminosity and contrast (all series were performed under automatic brightness mode as a feature of the HRTII-RCM); the current results were obtained without any pre-rocessing of the images. Real-time mapping of the SNP was achieved in all five healthy volunteers and showed the presence of epithelial ACM ridges with K-structures directly underneath in all cases (Fig. 1A). Moreover, the occurrence of this phenomenon was seen to correlate directly with the development of a superficial ACM pattern, as confirmed by slit lamp (Fig. 1B). Our results therefore verify the findings of Kobayashi et al. ¹⁰ that K-structures are located beneath the epithelial ACM ridges. 

Image Registration

Several image sequences were analyzed. Calculation of subimage shift vectors and reduction of motion-induced distortion effects showed good results in most of the images. However, approximately 2% to 3% of the images were extremely distorted because of fast motion (Fig. 2B). In these cases, subimage registration did not yield reliable results and the shift values for these subimages were then interpolated from the preceding and following subimages. In a few cases, this had to be performed manually to produce satisfactory results. Figures 6A–C shows representative images at different depths from an image sequence passing through the SNP. Figures 6D–F show the corresponding motion-corrected images. 

Figure 6.

 

Representative in vivo confocal microscopic images of cornea adjacent to the SNP layer with subsequent image registration. (A) Level of epithelial basal cells: K-structures are visible (arrows), and structures of the SNP are present (arrowheads) at a depth of 46 μm. (B) Level of SNP: K-structures are visible (arrows), and anterior stroma is visible through the K-structures, at a depth of 60 μm. (C) Level of anterior stroma (as) and Bowman's membrane (bm), depth 68 μm. (D-F) Registration procedure applied to corresponding images A-C. The motion-induced distortions in these images were in the order of magnitude of 10 pixels (∼10 μm) (i.e., several times the width of corneal nerve fibers <5 μm]). Note that the vertical sides of the depicted images show the nonlinear progression of motion-induced distortion effects.

Volume Reconstruction

Section images (Figs. 7B, 7C) through the reconstructed volume (Fig. 7A) clearly show the formation of ACM ridges at the level of the basal epithelium and Bowman's membrane. 

Figure 7.

 

Three-dimensional (3D) reconstructed volume and corresponding section images. (A) 3D reconstructed volume, total image height: 60 μm (30 slices). (B, C) Sections along the x-z–plane (B) and y-z–plane (C), showing the ridge-like deformations of the ACM (arrows). (D) Section image B, smoothed and with automatic tracing of SNP layer. (E) Section image C, smoothed and with automatic tracing of SNP layer.

Nerve Fiber Layer Extraction

Processing of two dimensional image series permitted reconstruction of the SNP level free of ACM ridges. The tracing of the bright SNP layer in section images is shown in Figs. 7D and 7E. All section tracings together form the depth map (Figs. 3A, 3B, 4A, 4B, 5A, 5B) which illustrates depth below the corneal surface of the SNP. The depth map allows the image at the level of the SNP to be assembled from the voxel data (Figs. 3C, 4C, 5C). The quality of the reconstructed image of the SNP layer depends heavily on accurate registration results. Moreover, the tracing process is sometimes led away from the SNP and into the bordering stroma region by hyperreflective keratocyte nuclei. The averaged elevation of ACM ridges was 20.6 μm (range, 8.7–34.0 μm); consequently, the maximum range of 60 μm with the internal scan mode of the HRTII was sufficient to depict the SNP structures of interest. 

The total processing time for each image sequence was approximately 3 minutes. About half of that time was used for registration (including volume reconstruction) and the other half for nerve fiber layer extraction. 

Discussion

Although a number of research studies have been conducted using confocal microscopy in corneal investigations, only a few approaches to volume reconstruction from in vivo confocal images have been published. Early works simply stacked up the images, ignoring any motion-induced shift or distortion effects. ^{23 –25} While this method has proved adequate for measuring total corneal thickness or corneal sublayer thickness, it generally does not permit further qualitative or quantitative assessment of corneal structures. Recently, Scarpa et al. ²⁶ described a customized algorithm using normalized image correlation to calculate the shift along the x- and y-axes between consecutive images. The authors used a slit scanning system (ConfoScan 4; Nidek Technologies, Padua, Italy) and, apart from the need to exclude approximately 3% of the images, they reported satisfactory registration results. The image sequences contained up to 144 images with interslice distances varying between 1 μm and 6 μm. The major drawback with their proposed method is its inability, because of its rigid registration approach, to handle motion-induced distortion effects inherent in the images. This may not be necessary for the evaluation of every type of corneal structure, but the greater the requirement for high-quality registration, the more attention has to be paid to this phenomenon that is inherent in every scanning image acquisition process. 

When comparing the axial resolutions (the thickness of the volume slice that contributes to a single image) offered by competing types of confocal microscopes (tandem scanning, 9 μm ²³ ; slit scanning, 25 μm ²⁶ ; and laser scanning, 7–8 μm ²⁷ ), CLSM affords an excellent opportunity to render tissue in authentic proportions. Two things have to be taken into consideration when imaging the SNP with a two-dimensional (2D) imaging technique such as confocal microscopy: the process of acquiring 2D images from three-dimensional (3D) structures such as nerve fibers and the spatial arrangement of the SNP as a whole. The spatial arrangement of the SNP does in fact support 2D imaging techniques. After penetrating Bowman's membrane perpendicularly, the corneal nerve fibers bend sharply at a right angle and then run tightly along the basal membrane of the corneal epithelium. In the absence of ACM deformations, the epithelial basal membrane can be described as almost perfectly parallel to the corneal surface. As long as the microscope is flatly attached to the corneal surface, the depiction of the SNP is possible with 2D image acquisition techniques. Different reconstruction techniques (volume rendering [400 × 400 × 60 μm], cross-section, en face view, oblique section, and surface reconstruction) using CLSM and commercially available software for image registration (Amira 3.1; TGS Inc, San Diego, CA) have been performed and published. ^27,28 Nevertheless, as with Scarpa's approach, only rigid image registration was performed, and consequently image distortions could not be handled. Because of motion artifacts, approximately 50% of all data were not suitable for additional reconstruction. 

To date, all published approaches to volume reconstruction from in vivo confocal microscopy (IVCM) of the cornea have used either no image registration or rigid image registration only. The method proposed here uses nonrigid, elastic image registration for the task. The algorithm takes account of the special image acquisition process used in confocal scanning microscopy. The results of this pilot study have shown that this approach is capable of reducing motion-induced distortion effects characteristically encountered with IVCM and that it promises superior registration results. Nevertheless, several restrictions and problems remain in its current implementation. More work is required to guarantee reliable registration results in all images, including those taken during fast eye motion. Even though the processing time of approximately 1.5 minutes (for registration and volume reconstruction) is suitable for use in clinical practice, the need for manual intervention and correction, however rare, is not. 

To assess the subbasal nerve fibers over the whole image area despite the presence of ACM ridges, we have developed a new method incorporating digital image processing techniques. The reconstructed image shows the results of subbasal nerve layer extraction with clearly visible continuous subbasal nerve fibers (Figs. 3C, 4C, 5C). Additional research needs to be invested in this processing step to eliminate the reported erroneous tracings into the bordering stroma. 

In conclusion, one major problem inherent in the evaluation of corneal nerve fibers is that it is often not possible to move the SNP into the focal plane over the entire image area. This is because of ridge-like deformation of the SNP region in connection with the formation of ACM. Using a novel procedure for image acquisition and processing, we are able to detect these features and reconstruct an image of the SNP. In the five subjects examined, the ridges had an average measured elevation of approximately 20 μm. 

In this article, we present preliminary results with a novel method for significantly reducing the influence of ACM on confocal microscopic imaging of the SNP. As already alluded to above, it is now necessary to further refine the proposed image processing methodology to provide a fast, robust, automatic software module. In addition, our preliminary findings require confirmation in more extensive studies. Equally important is the development of automatic nerve fiber recognition software that is optimized for the statistically reliable morphometric characterization of corneal nerve patterns. This will allow for the assessment of data for prospective and multicenter studies in the normal cornea and in corneal pathology. Possible future work might also include improvements in the imaging process of the SNP itself, such as using online image processing results to control the internal or external microscope focus drive to adaptively stay at the depth of the SNP or, in the case of ACM, to continuously scan between the top and bottom of ACM ridges to automatically acquire the necessary images for reconstruction. 

Footnotes

 Supported in part by the Deutsche Forschungsgemeinschaft (Transregio 37, Micro- and Nanosystems in Medicine—Reconstruction of Biological Functions).

Footnotes

 Disclosure: S. Allgeier, None; A. Zhivov, None; F. Eberle, None; B. Koehler, None; S. Maier, None; G. Bretthauer, None; R.F. Guthoff, None; O. Stachs, None.