2.1.

Noncontact, 3D Large Field-of-View Corneal Imaging with Gabor-Domain Optical Coherence Microscopy

GDOCM was first developed for contact imaging, which is preferred in applications that benefit from index-matching between the microscope and the sample, such as dermatology.^27,39,40 The contact imaging modality, however, is not desirable for corneal imaging, since physical contact may cause damage to the tissue, and requires exposing the tissue to the outside environment. The GDOCM microscope was recently redesigned to support two imaging modalities: contact and noncontact with 15-mm working distance. The GDOCM system shown in Fig. 1(a) can be used to image corneal tissue stored in viewing chambers using a noncontact imaging modality, as shown in Fig. 1(b). For corneal tissue imaging with both GDOCM and SM, the PMMA viewing chamber causes losses in the signal acquired due to the specular reflection at the air/PMMA interface, and due to the scattering and absorption inside the material; in the case of GDOCM, high signal-to-noise ratio images of the corneal tissue can still be achieved due to the capability to dynamically focus the beam on the tissue below the lid of the container.

Fig. 1

(a) GDOCM system and (b) GDOCM microscope operating at 15-mm working distance to image corneal tissue through a viewing chamber. There is an air gap between the microscope and the viewing chamber.

Data from six corneas previously imaged with GDOCM with contact modality (180 images total) and manually counted³² were used to train the CNN for the automatic segmentation of endothelial cells. Five locations on each of six corneas were imaged; six repeated imaged were taken at the same location after repositioning the sample under the microscope, for a total of 180 images. The purpose of repeating the imaging at the same location after repositioning the sample under the microscope was to augment the dataset while having a limited number of corneas available, and to acquire multiple images affected by real imaging issues, such as noise and other imaging artifacts potentially introduced by the operator when positioning the probe.

When introducing a new imaging modality, such as GDOCM, validation against the gold standard (SM in this case) is necessary to demonstrate noninferiority of the new solution. To assess the performance of the automated cell counting procedure, 10 donor corneas were imaged with SM and with GDOCM in noncontact modality through a viewing chamber (endothelium side up). Optisol GS (Bausch + Lomb, Rochester, New York) was used as the storage medium. The GDOCM imaging configuration is shown in Fig. 1(b). The donor tissue was warmed at room temperature for 2 h prior to imaging.

An example of the 3D high-definition, wide field-of-view imaging capability of GDOCM is shown in Fig. 2.

Fig. 2

GDOCM images of an ex vivo human cornea (field of view $1{\mathrm{mm}}^2$) imaged through the viewing chamber. (a) 3D view. (b) Full-depth cross-sectional image showing the corneal layers: EP, epithelium; BL, Bowman’s layer; ST, stroma; DM, Descemet’s membrane; EN, endothelium. (c) En face view of the epithelium. En face images of the (d) middle and (e) posterior stroma reveal stromal keratocytes (white arrows). (f) En face view of the transition between stroma and endothelium, with endothelial cells visible. The bars are $100\mu \mathrm{m}$.

The 3D corneal images acquired by GDOCM, such as SM images, are affected by the curvature of the endothelium. As a consequence, it is not possible to visualize the endothelial cell layer in a single en face image, as shown by Fig. 3.

Fig. 3

En face views at the transition between stroma and endothelium. (a)–(d) Separated in depth by $3\mu \mathrm{m}$ each. Bars: $100\mu \mathrm{m}$. The hyper-reflective endothelial cells appear brightest when in focus.

Each GDOCM en face view is separated by $\sim 1.5\mu \mathrm{m}$, which corresponds to the axial sectioning capability of the system. The en face images shown in Figs. 3(a)–3(d) span a total depth of $9\mu \mathrm{m}$ and are each separated by $\sim 3\mu \mathrm{m}$; it can be seen that only a fraction of the total endothelial cells is in focus and clearly visible in each image. The hyper-reflective endothelial cells appear the brightest when they are in focus. In addition, cells at the center of the cornea are brighter than at the periphery, since there is some loss of light at the periphery due to the natural curvature of the cornea causing some of the backscattered light not to be collected by the microscope. As such, the effective area considered in the analysis was $0.2{\mathrm{mm}}^2$, corresponding to the central portion of the cornea with a diameter of 0.5 mm. It should be noted that this constitutes a significant improvement (6 to $12\times$) in the number of cells counted over the current practice, which consists of eye bank technicians manually counting only a total of 50 to 100 cells within the $0.12{\mathrm{mm}}^2$ area imaged with SM. This improvement is due to two concurrent factors: (1) the area of analysis is increased with GDOCM compared to the state-of-the art SM instruments, and (2) with automated cell counting all cells in the area of analysis are assessed, while the manual counting performed by eye bank technicians is done by assessing only 50 to 100 cells. With a typical ECD of $3000{\text{cells}/\mathrm{mm}}^2$, the number of cells present in a $0.2\text{-}{\mathrm{mm}}^2$ area is 600, which is 6 to 12 times the number of cells manually assessed (50 to 100).

2.2.

Automatic Segmentation and Numerical Flattening of the Endothelium

Numerical methods can be implemented to digitally flatten the endothelium in a 3D GDOCM image. Previous flattening attempts of GDOCM images consisted of detecting the peak in each A-scan (1D depth profile) and repeating this step to the entire volumetric image to extract the endothelial surface; after polynomial fitting, the pixels of the 3D image were shifted to produce a flattened view of the endothelium.^31,32 The drawback of this approach is that artifacts are often present in the flattened image due to the 1D approach of this method. Here, we propose the 3D approach illustrated in Fig. 4, which consists of the following steps. Step 1: Gaussian blur for noise reduction. Gaussian blur, a form of low-pass filter, is a standard noise reduction technique used in image processing to reduce the sharpness images, thus reducing noise artifacts from the acquired image. Step 2: Segmentation of the cornea using Huang binarization.⁴¹ Ideally, the acquired volume should be bimodal, i.e., have bright voxels over the entire cornea and dark voxels elsewhere. In practice, the volume is still noisy, even after denoising. The Huang binarization attempts to find the optimal binarization threshold by minimizing the fuzziness of the image. Step 3: Median filtering to reduce binarization artifacts. Binarization can increase the noise in the image, especially the so-called salt-and-pepper noise. Median filtering is a well-established image processing technique to reduce this type of noise.⁴² Step 4: Identification of the endothelial surface with RANSAC fitting.⁴³ Step 5: Max-intensity projection of a slice below the fitted surface to account for estimation errors in steps 2 and 3. The outcome of the 3D flattening approach is an artifact-free en face view of the endothelium.

Fig. 4

Automated segmentation of the endothelium from a GDOCM image shown here in a 2D cross-section. (a) Original image. (b) After Gaussian blur (step 1). (c) After Huang binarization (step 2). (d) After median filtering (step 3). (e), (f) Segmented endothelial surface in green (e) before and (f) after RANSAC fitting (step 4). The bars are $100\mu \mathrm{m}$.

2.3.

Automated Cell Segmentation and Counting with Machine Learning

A dataset with 180 GDOCM images (obtained from imaging six corneas at five locations each), and the corresponding manual tracing of the endothelial cells over the central 0.5-mm diameter of the flattened endothelium was used for training and testing a CNN–VGG-16 (see Fig. 5).⁴⁴ We applied a transfer learning approach using the pretrained VGG-16 network. The network is pretrained on more than a million images from the ImageNet database. The pretrained model has already learned to extract features from different image categories and can improve the classification results even with small training datasets. Data augmentation was also used by applying randomly distributed translations of $\pm 10\text{pixels}$ in both directions, specular reflections and rotations. Scaling and skewing were not applied to preserve the morphology and the sizes of the cells. We trained the CNN six times using a leave-one-out cross-validation strategy.⁴⁵ The training was done on five manually traced corneas (150 images) and tested on the sixth cornea (30 images); this procedure was repeated six times, leaving out a different cornea for testing each time.

Fig. 5

VGG-16 architecture⁴⁴ used for automated endothelial cell segmentation. VGG-16 is a type of CNN. CNNs, also called shift invariant artificial neural networks, are a type of artificial neural network` specially designed for image processing and modeled after the visual cortex of animals. The convolutional layers (denoted by the prefix Conv in the figure) mimic the functions of visual receptors, while the pooling layers filter the most salient activations from the lower layers. Dense layers are regular fully connected multilayer perceptrons, used for classification.

Each pixel of the images in the dataset was annotated as belonging to one out of three classes: cell border, cell interior, and noise. For the training, we used stochastic gradient descent with a momentum of 0.9, an initial learning rate of 1e-3, an L2 regularization of 5e-4, a minibatch size of 16 images, and a maximum of 400 epochs.

To obtain the segmentation of the cells in the testing step, the raw classification results from the trained network were postprocessed as follows: a binary mask was extracted from the pixels classified as cell border. A two-pixel dilation, to close small gaps in the borders, followed by skeletonization, was applied to the binary mask. Regions smaller than 100 square pixels or larger than 750 square pixels were excluded from the segmentation as representing oversegmentation or noisy areas; the chosen interval included 95% of the cells in the ground truth. An example of the fully automated procedure for ECD analysis from the en face view of the flattened endothelium is shown in Fig. 6. The analysis was performed on the central cornea with 0.5-mm diameter to match the manually annotated data available.

Fig. 6

Automated cell counting procedure. (a) After automatic segmentation of the endothelium and numerical flattening, (b) the region with 0.5-mm diameter in the center of the cornea is selected for analysis and cropped, to match the manually annotated data available. (c) The CNN provides automatic segmentation from the image. (d) After skeletonization of the segmented borders, (e) the cell areas can be quantified.

2.4.

Comparison with Specular Microscopy

For comparison with the gold standard of SM, 10 corneas were analyzed at the eye bank using a Konan CellChek D+ SM with the standard evaluation process, in which a technician manually selects 50 to 100 cells and the ECD is obtained via a center/flex center method. The imaging with GDOCM was conducted on the corneas stored in the same viewing chambers as for SM evaluation, and on the same central area of the cornea. After numerical flattening of the endothelium, a CNN was used to automatically obtain the ECD.

3.1.

Endothelium Segmentation and Flattening

The automated 3D endothelium segmentation and flattening procedure was applied to 3D GDOCM images to produce en face views of the endothelium. A comparison with the 1D approach is shown in Fig. 7.

Fig. 7

En face view of the flattened endothelium. (a) A 1D approach introduces artifacts that appear as lines creating discontinuities in the image, while the proposed 3D approach results in (b) artifact-free flattening.

While the 1D approach can result in artifacts [as shown in Fig. 7(a)] due to mismatch between adjacent points, the 3D approach results in an artifact-free en face view of the endothelium [see Fig. 7(b)]. The flattened endothelia for the 10 corneas imaged with noncontact imaging modality through the viewing chambers are shown in Fig. 8.

Fig. 8

Flattened endothelium for 10 corneas imaged with GDOCM using the noncontact imaging configuration of Fig. 1(b). Dark bands caused by folding of the tissue are visible in a few images.

The darker periphery in the images is caused by the lower scattering signal being collected from regions where the tissue is less normal to the microscope than the center of the cornea.

3.2.

Automated Cell Segmentation

The manually annotated dataset available consisted of 180 images from six corneal samples imaged with GDOCM.³² The CNN was trained using manually traced images from five corneal samples, for a total of 150 images (six repeated images at five different locations on each cornea). Postprocessing of the identified borders yielded the cell segmentation. The performance of the CNN was tested on 30 images from the remaining cornea (six repeated images at five different locations). This procedure was repeated six times in a leave-one-out cross-validation strategy. An example of the results obtained is shown in Fig. 9 overlaid with the manual annotation.

Fig. 9

Cell counting results over a region at the center of the cornea with 0.5-mm diameter. The cells colored in white were found by both manual and automated procedures; the cells in teal were found only by the automated procedure; and the cells in red were only found by the manual procedure. In some cases, the automated procedure identified $\sim 30\%$ more cells than the manual annotation.

An average of 316 cells per image was counted from the GDOCM images with the automated procedure, compared to 307 with the manual procedure. The comparison between manual and automated cell counting on cell density values for the corneas used for assessing the CNN performance is shown in Fig. 10. Only cells identified by both methods have been included in this analysis. We chose to only include cells found by both methods, because in reviewing the ground truth we noticed that several instances of cells found by the CNN that had incorrectly not been annotated in the manual analysis. On average, the CNN found 12% more cells than what was in the manual annotation.

Fig. 10

Comparison of cell counting results for manual (black line) and automated (blue line) procedures on six corneas (cornea 1 to cornea 6) imaged at five locations each (L1-L5), with six repeated images at each location (total 180 images). The automated cell counting results were obtained with a leave-one-out cross validation. Error bars indicate the standard deviation for the six repeated measurements at each location.

The average difference in ECD between the two modalities was 0.3%, and their Pearson correlation had an $R^2$ of 0.94 ($p<0.001$), as shown in Fig. 11(a).

Fig. 11

(a) Comparison of manual and automated ECD estimation from GDOCM images for six corneas (C1 to C6), each imaged at five locations with six repetitions (30 images per cornea). (b) Bland–Altman plot showing show the difference between ECD measurements by the automated and manual methods plotted against the mean ECD.

The difference in mean ECD between the two automated and manual cell count was $22\text{cells}/{\mathrm{mm}}^2$, and Bland–Altman analysis revealed that 95% of the differences in ECD were within $\pm 185\text{cells}/{\mathrm{mm}}^2$ of the mean difference, as shown in Fig. 11(b). The Bland Altman 95% confidence interval includes zero.

In some cases, the automated procedure resulted in undersegmentation (see Fig. 12), thus causing lower ECD results.

Fig. 12

Example of undersegmentation with the automated cell counting method. (a) GDOCM image of the endothelial cells. (b) Manual annotation of the cell borders. (c) Outcome of the CNN, with two cells incorrectly identified as a single cell. (d) Overlay of (a)–(c).

The average Dice score⁴⁶ was 0.66 when computed on all cells, and 0.77 when computed only on ground truth cells (thus excluding false positives due to incomplete ground truth).

3.3.

Comparison with Specular Microscopy

To assess the GDOCM automatic cell counting relative to SM, 10 additional corneas were imaged for which we had a complete report from the eye bank, including ECD assessed on a Konan CellChek $\mathrm{D}+$ SM. The tissue data are summarized in Table 2. The time between death and tissue preservation was on average 11.1 h (range: 4.5 to 18.6). The GDOCM imaging took place an average of 5.6 days after the SM imaging (range: 2.1 to 7.2).

Table 2

Donor age, time between death and cooling, time between death and preservation, time between preservation and SM imaging, and time between SM imaging and GDOCM imaging for 10 corneas imaged in this study.

The GDOCM automated cell count results were conducted on the central 0.5 mm of each image. The results are shown in Fig. 13.

Fig. 13

Comparison of (left) SM images, (middle) GDOCM central 0.5-mm diameter images, and (right) automatic segmentation for the 10 corneas of Table 2. Images have the same scale.

The automated cell count results were manually corrected for the cases in which two or more cells were combined in one. An average of 180 cells were counted from the GDOCM images with the automated procedure. The regions imaged with the two modalities were near the center of the cornea but not identical. Also, it should be noted that the SM results likely suffer from selection bias (i.e., the operator only selects a subset of 50 to 100 cells within the field of view). These factors can account for differences in the counts. The results are shown in Table 3 and Fig. 14.

Table 3

Cell counting from SM images (semiautomated with manual cell selection) and GDOCM images (automated) for the ten corneas imaged in this study.

Fig. 14

Comparison of semiautomated ECD estimation from SM images and automated ECD estimation from GDOCM images for 10 corneas of Table 2.

The average difference in ECD between the two modalities was 3.7%, and their Pearson correlation had an $R^2$ of 0.91 ($p<0.001$), as shown in Fig. 15(a).

Fig. 15

(a) Comparison of semiautomated ECD estimation from SM images and automated ECD estimation from GDOCM images for the 10 corneas of Table 2. (b) Bland–Altman plot showing show the difference between ECD measurements by SM and GDOCM plotted against the mean ECD.

The difference in mean ECD between the two modalities was $99{\text{cells}/\mathrm{mm}}^2$, and Bland–Altman analysis revealed that 95% of the differences in ECD were within $\pm 268{\text{cells}/\mathrm{mm}}^2$ of the mean difference, as shown in Fig. 15(b). The Bland–Altman 95% confidence interval includes zero.

The entire pipeline for automated cell counting from GDOCM images, segmentation, postprocessing, and counting, takes an average of 2 s per cornea on a common laptop (a MacBook Pro with a 3.1 GHz Intel Core i7 processor and 8 GB of memory). The memory footprint of the trained network is about 122 MB. By comparison, a trained technician takes on average 20 to 30 min to perform SM evaluation.