Open Access
Artificial Intelligence  |   February 2025
RetOCTNet: Deep Learning–Based Segmentation of OCT Images Following Retinal Ganglion Cell Injury
Author Affiliations & Notes
  • Gabriela Sanchez-Rodriguez
    Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
    Department of Ophthalmology, Emory University, Atlanta, GA, USA
    Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, GA, USA
  • Linjiang Lou
    Department of Ophthalmology, Emory University, Atlanta, GA, USA
  • Machelle T. Pardue
    Department of Ophthalmology, Emory University, Atlanta, GA, USA
    Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, GA, USA
    Centre for Visual and Neurocognitive Rehabilitation, Atlanta VA Healthcare System, Atlanta, GA, USA
  • Andrew J. Feola
    Department of Ophthalmology, Emory University, Atlanta, GA, USA
    Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, GA, USA
    Centre for Visual and Neurocognitive Rehabilitation, Atlanta VA Healthcare System, Atlanta, GA, USA
  • Correspondence: Andrew J. Feola, Department of Ophthalmology, Emory University, B2503, Clinic B Building, 1365B Clifton Road NE, Atlanta, GA 30322, USA. e-mail: [email protected] 
Translational Vision Science & Technology February 2025, Vol.14, 4. doi:https://doi.org/10.1167/tvst.14.2.4
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Gabriela Sanchez-Rodriguez, Linjiang Lou, Machelle T. Pardue, Andrew J. Feola; RetOCTNet: Deep Learning–Based Segmentation of OCT Images Following Retinal Ganglion Cell Injury. Trans. Vis. Sci. Tech. 2025;14(2):4. https://doi.org/10.1167/tvst.14.2.4.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose: We present RetOCTNet, a deep learning tool to segment the retinal nerve fiber layer (RNFL) and total retinal thickness automatically from optical coherence tomography (OCT) scans in rats following retinal ganglion cell (RGC) injury.

Methods: We created unilateral RGC injury by ocular hypertension (OHT) or optic nerve crush (ONC), and contralateral eyes were not injured. We manually segmented the RNFL and total retina of 3.0-mm radial OCT scans (1000 A-scans per B-scan, 20 frames per B-scan) as ground truth (n = 192 scans). We used these segmentations for training (80%), testing (10%), and validation (10%) to optimize the F1 score. To determine the generalizability of RetOCTNet, we tested it on volumetric scans of a separate cohort at baseline and 4, 8, and 12 weeks post-ONC.

Results: RetOCTNet's segmentations achieved high F1 scores relative to the ground-truth segmentations created by human annotators: 0.88 (RNFL) and 0.98 (retinal thickness) for control eyes, 0.84 and 0.98 for OHT eyes, and 0.78 and 0.96 for ONC eyes, respectively. On volumetric scans 12 weeks post-ONC, RetOCTNet calculated thinning of 29.49% and 10.82% in the RNFL and retina at the optic nerve head (ONH) and thinning of 38.34% and 9.85% in the RNFL and retina superior to the ONH.

Conclusions: RetOCTNet can segment the RNFL and total retinal thickness of both radial and volume OCT scans. RetOCTNet can be applied to longitudinally monitor RNFL in rodent models of RGC injury.

Translational Relevance: This tool automates RNFL and retinal thickness measurement for rat OCT scans following RGC injury, reducing analysis time and increasing the consistency between studies.

Introduction
Glaucoma is the leading cause of irreversible blindness worldwide, and it is projected to affect 111.8 million people by 2040.1,2 Glaucoma is an optic neuropathy characterized by the loss of peripheral vision stemming from the loss of retinal ganglion cells (RGCs) and their axons.3 Structural changes in glaucoma are often observed by fundus or optical coherence tomography (OCT) imaging and include increased cup-to-disc ratio, cupping of the optic disk, and thinning of the retinal nerve fiber layer (RNFL).4 
OCT is advantageous because it is a noninvasive technique that allows for in vivo longitudinal imaging of the retina and optic nerve head (ONH) in patients.57 This imaging approach has become a commonly used practice in glaucoma diagnosis, as its diagnostic capabilities help discriminate glaucomatous eyes based on RNFL thickness measurements.7 However, the manual analysis of retinal OCT images is a time-consuming task that is performed by expert annotators. This has led to an increase in the number of automatic OCT retinal analysis tools, further fueled by the number of public OCT data sets available.8 Currently, most of these approaches are developed for clinical applications and have limited applicability to experimental research models of glaucoma. The Iowa Reference Algorithms (Retinal Image Analysis Lab, Iowa Institute for Biomedical Imaging, Iowa City, IA, USA) has developed several tools for retinal image analysis for both human and murine eyes.9,10 These methods were initially based on features of the OCT without leveraging the power of deep learning (DL). 
DL is a subset of artificial intelligence based on deep neural networks, which has become increasingly used for glaucomatous assessment of OCT images. Quantitative measurements and thickness maps, among others, can be extracted from DL models.11 Most DL segmentation consists of identifying retinal layers and/or retinal lesions by solving pixel-wise classification.8 Pekala et al.12 proposed a novel method for the automatic segmentation of five retinal layers (preretinal space, nerve fiber layer, ganglion cell layer, inner plexiform layer, and another surface comprised from the inner nuclear layer until the choriocapillaris) in OCT images. This approach uses fully convolutional DenseNets with skip connections based on the encoder–decoder architecture proposed by Jégou et al.13 Using this same architecture, Sedai et al.14 proposed a semi-supervised approach with an uncertainty guided student–teacher network model that reduced the number of ground-truth images required. Others have focused on leveraging the anatomic features within OCT images of human patients from clinical data sets. For example, Lazaridis et al.15 combined attention mechanisms with constraints to the receptive field of the convolutional layers to improve the model's ability to extract features oriented in vertical or horizontal directions. He et al.16 focused on a multitask network for surface layer segmentation in individual A-scans of a B-scan to produce smooth, structured continuous layer surfaces without the explicit constraints that graph methods require. Viedma et al.17 proposed applying an instance segmentation architecture to OCT images, Mask R-CNN,18 instead of the more commonly used encoder–decoder backbone, considerably reducing inference time. 
While many of these DL models were developed on human data sets, several groups have developed DL approaches for experimental research models of retinal injury. Antony et al.19 developed ASiMOV, a graph-based algorithm that combines machine learning to segment 10 retinal surfaces in healthy mouse eyes and 6 in mouse eyes with light-induced retinal damage. Mishra et al.20 used dynamic programming optimization to find optimal layer boundaries in rodent retinas. Srinivasan et al.21 presented a method based on graph theory and dynamic programming combined with a support vector machine algorithm trained on wild-type mice and rhodopsin knockout mice. Morales et al.22 presented a DL-based approach to segment rodent eyes based on the ReLayNet network, which was able to accurately segment six retinal layers of control and endothelin-1–injected eyes, a drug producing vasoconstriction and ischemia. Despite the applications of these DL algorithms to preclinical data sets, none of these algorithms have been developed with data sets that included OCT scans from eyes after RGC injury. This poses a limitation to the use of these algorithms to monitor RNFL thickness in experimental models of glaucoma or RGC injury. Here, we present RetOCTNet, a fully automated DL approach to segment the RNFL and total retinal thickness in rat eyes, trained on healthy eyes and eyes after RGC injury with different levels of retinal damage. 
Methods
Data Collection, Segmentation, and Preprocessing
We created unilateral RGC injury by ocular hypertension (OHT) (n = 38, 5- to 12-month-old Brown Norway rats) or optic nerve crush (ONC) (n = 12, 11- to 12-month-old Long-Evans rats) with the contralateral (CL) eyes serving as an internal reference. We acquired four evenly spaced 3.0-mm radial OCT scans (1000 A-scans per B-scan, registering and averaging 20 frames per B-scan; Bioptigen 4300; Leica Microsystems, Buffalo Grove, IL, USA) centered at the ONH. OCT images (2024 × 1000 pixels) had an axial resolution of 2.75 µm and a lateral resolution of 3 µm. Data used for development were originally collected in previous studies.23,24 In brief, we acquired OCT scans from the OHT cohort at 0 (baseline), 4, and 8 weeks. For the ONC cohort, we acquired OCT scans at 12 weeks after injury using the same radial scan parameters for the OHT cohort. We treated the four B-scans of each eye at each time point as independent samples. We divided the data randomly into training (80%), testing (10%), and validation (10%) data sets. The division of eyes is shown in Supplementary Table S1. We ensured scans from the same eye were only present in one data set (train, test, or validation) and that the model was trained on all possible types of injury. Additionally, the number of controls and injured eyes was evenly distributed in the training set. An expert human annotator created multiclass labels for each of these scans. The masks labeled the background, RNFL, and the space between the internal limiting membrane to the choroid (Fig. 1 and additional layers labeled in Supplementary Fig. S1). We normalized the intensity of the averaged OCT scans across each data set independently to a unit norm and zero mean. To reduce computational time, we took the original image (1024 × 1000) and discarded the bottom of images and masks as they only contained additional background information. This resulted in an image size of 512 × 1000 pixels. We padded these images (i.e., added a border) to make them 512 × 1024-pixel images. We then evenly divided the image into two 512 × 512 nonoverlapping subimages. Similar to previous,2527 we refer to these subregions as patches. A walkthrough of the patching process can be found in Supplementary Figure S8
Figure 1.
 
Example of RetOCTNet and human segmentations overlaid on a control eye, an optic nerve crush eye 12 weeks postinjury, and an ocular hypertensive eye 8 weeks postinjury.
Figure 1.
 
Example of RetOCTNet and human segmentations overlaid on a control eye, an optic nerve crush eye 12 weeks postinjury, and an ocular hypertensive eye 8 weeks postinjury.
Model Design and Configuration
The RetOCTNet architecture directly adopts the architecture proposed by Jégou et al.13 without any modifications. We utilized a publicly available Keras implementation of the Keras-contrib GitHub repository.28 For the visualization of this architecture, we refer the reader to the original publication by Jégou et al.13 
We tested different hyperparameter configurations using Wandb,29 a framework-agnostic experiment tracking tool for machine learning. We defined an initial grid or range where the hyperparameter values should remain providing a bounded parameter search space. We used the Bayesian Optimization option. For more information on Bayesian Optimization, we refer the reader to various studies.3032 The list of variables, parameter space, and final optimized hyperparameters are in Supplementary Table S2
We used the Adam33 optimizer and monitored cross-entropy loss and accuracy metrics during training to assess its learning progress. 
Once trained, we used RetOCTNet to segment the RNFL and total retina (Fig. 1). To evaluate the quality of the RetOCTNet segmentations compared to the ground truths, we stitched the segmented patches together. Since the patches came from nonoverlapping regions of the same image, the stitching involved placing these patches adjacent to each other in their original spatial arrangement, reconstructing the full segmented image from its individual patches and subsequently removing the padding. We computed F1 scores (equivalent to the Dice coefficient in segmentation, or binary, data sets34,35), accuracy, intersection over union (IOU), precision, and recall (Supplementary Table S3). We report these values for the RNFL layer and total retinal thickness (inner limiting membrane to the choroid). To better interpret these metrics and appraise agreement with our desired outcomes (e.g., RNFL and total retinal thickness values), we evaluated the thickness measurements for the RNFL and total retina by examining the correlation and Bland–Altman plots between RetOCTNet and ground truth. To assess whether the RetOCTNet and ground-truth thickness measurements were different, we used a two-sample Kolmogorov–Smirnov test (P < 0.05). 
Uncertainty Maps
We computed uncertainty maps to visualize regions segmented with less confidence (more uncertainty) and gain insight into the model performance of RetOCTNet predictions using the following formula:  
\begin{eqnarray}U\left( {\hat{y}} \right) = {\rm{\ }} - {\rm{\ }}\mathop \sum \limits_{c = 1}^C {{\hat{y}}_c} \times \log \left( {{{{\hat{y}}}_c}} \right)\end{eqnarray}
(1)
 
Similar to a previous approach,14 we calculated the segmentation uncertainty (\(U( {\hat{y}} )\)) for each pixel by computing the entropy of the softmax output produced by the network (\(\hat{y}\)). This is done in a classwise fashion (C). A prediction has high uncertainty when the network outputs similar probabilities for each class. We averaged the uncertainty for each injury type across all control, OHT, and ONC images within each test, train, and validation data set. 
Performance of RetOCTNet on a Different Data Set
To assess the generalizability of RetOCTNet, we tested it on a separate cohort of Long-Evans rats (n = 14, 11–12 months). In this cohort, we induced unilateral RGC injury by ONC and again used the contralateral eyes as internal reference (CL). We took OCT scans before injury (baseline) and at 4, 8, and 12 weeks post-ONC. OCT scans consisted of 2-mm × 2-mm volumetric OCT images (1000 A-scans per B-scan, registering and averaging 8 frames per B-scan, 29 B-scans per volume; Bioptigen 4300; Leica Microsystems). Relative to the scans used for training and developing RetOCTNet, these OCT scans altered the scan type (volume vs. radial), scan size (2 mm vs. 3 mm), and frames per B-scan (8 frames vs. 20 frames). We chose eight frames per B-scan because a preliminary study found that eight frames per B-scan was the minimum number of B-scan frames to average, which optimized the trade-off between noise reduction and file size (Supplementary Fig. S2) on our OCT system. 
After OCT scans were collected, we used RetOCTNet to compute RNFL and total retinal thickness of the ONC and CL eyes at each time point. We plotted the assessed RNFL and total retinal thickness at B-scans in the OCT volume: (1) at the ONH and (2) superior to the ONH (B-scans 15/29 and 21/29, respectively). This allowed us to assess if RetOCTNet required the presence of the ONH, similar to OCT images used during training, to accurately segment the retina from OCT scans. To compare if the RNFL and total retinal thickness changed after ONC, we performed a two-way ANOVA with eye (CL vs. ONC) and time (0, 4, 8, and 12 weeks) as our factors with a Šidák post hoc test. Statistics were performed in Prism (GraphPad Version 10; GraphPad Software, La Jolla, CA, USA) with a P < 0.05. 
Results
RetOCTNet Performance Compared to Human Annotators
We found that RetOCTNet scored 0.87 or higher for F1, precision, and recall metrics for the RNFL and total retinal thickness segmentation in the test data set (Table). Figure 2 shows examples of RetOCTNet segmentations. The correlation plots (Fig. 3) showed strong agreement between ground-truth thickness values and total retinal thickness, with R2 values of 0.91 in the training set and 0.97 in the test set. For RNFL thickness, the R2 values were 0.86 in the training set and more moderate in the test set at 0.71. To identify the bias between RetOCTNet and ground-truth thickness measurements, we examined Bland–Altman plots (Fig. 4). We found a relatively small difference in the RetOCTNet and ground-truth segmentations (maximum bias <1.7 µm). Next, we did not find a significant difference in the RNFL (P = 0.85) and retinal (P = 0.52) thickness measurements between RetOCTNet and ground-truth segmentations using a two-sample Kolmogorov–Smirnov test. 
Table.
 
RetOCTNet Performance Metrics Across Train, Test, and Validation Sets
Table.
 
RetOCTNet Performance Metrics Across Train, Test, and Validation Sets
Figure 2.
 
Example of RetOCTNet segmentation on OHT samples at 4 weeks post-OHT. Ground truth represents the manually annotated mask, and the prediction map shows the segmentation performed by RetOCTNet. Examples from the training set (two left columns) and test set (two right columns).
Figure 2.
 
Example of RetOCTNet segmentation on OHT samples at 4 weeks post-OHT. Ground truth represents the manually annotated mask, and the prediction map shows the segmentation performed by RetOCTNet. Examples from the training set (two left columns) and test set (two right columns).
Figure 3.
 
Correlation of the RNFL (left column) and total retinal (right column) thickness from RetOCTNet annotations versus ground truth (human annotations) for the training and test data set. The black dashed line is the unity line, the solid purple line represents the least squares solution to the best fit, and the hollow circles represent the individual thickness measurements.
Figure 3.
 
Correlation of the RNFL (left column) and total retinal (right column) thickness from RetOCTNet annotations versus ground truth (human annotations) for the training and test data set. The black dashed line is the unity line, the solid purple line represents the least squares solution to the best fit, and the hollow circles represent the individual thickness measurements.
Figure 4.
 
Bland–Altman plots from RNFL (left column) and total retinal (right column) thickness measurements of RetOCTNet and ground-truth annotations from the training and test data set. The purple solid lines represent the bias while the dashed lines are the 95% limits of agreement.
Figure 4.
 
Bland–Altman plots from RNFL (left column) and total retinal (right column) thickness measurements of RetOCTNet and ground-truth annotations from the training and test data set. The purple solid lines represent the bias while the dashed lines are the 95% limits of agreement.
Uncertainty
We report the average uncertainty of RetOCTNet across different RGC-injury models in the validation data set (Fig. 5). Control animals have consistently less uncertainty, indicating it is easier for RetOCTNet to classify the pixels in the RNFL and retina from uninjured eyes. Uncertainty increases as RGC damage advances. In general, much of the uncertainty was located near the ONH and retinal boundaries. 
Figure 5.
 
Uncertainty projection in the validation data set. The ONH region is the region with the largest uncertainty values.
Figure 5.
 
Uncertainty projection in the validation data set. The ONH region is the region with the largest uncertainty values.
RetOCTNet Performance on a Different Data Set
To test the applicability of RetOCTNet on a different type of OCT scan, we took the fully trained algorithm and applied it to a data set with a different scan type (volume vs. radial) and higher noise due to fewer frames being averaged (8 frames compared to 20 frames per B-scan). RetOCTNet was able to segment the RNFL and total retina in the volume scans with and without the presence of the ONH (Fig. 6). RetOCTNet also found thinning of the RNFL and total retina following ONC (Fig. 7). At the ONH, the average RNFL thickness decreased by 16% (P = 0.016) to 29% (P < 0.0001) from 4 weeks to 12 weeks after ONC injury. The total retinal thickness decreased by 5% (P = 0.014) and 10% (P < 0.0001) at 4 weeks and 12 weeks after ONC. We found a similar trend superior to the ONH. In this region, the RNFL and retinal thickness decreased by 27% (P < 0.0001) and 9% (P < 0.0001) 12 weeks after ONC compared to the CL eyes. 
Figure 6.
 
Samples of OCT scans and masks produced by RetOCTNet from OCT volumes at two different locations: superior to the ONH (scan: 21/29) (top) and at the ONH (scan 15/29) (bottom).
Figure 6.
 
Samples of OCT scans and masks produced by RetOCTNet from OCT volumes at two different locations: superior to the ONH (scan: 21/29) (top) and at the ONH (scan 15/29) (bottom).
Figure 7.
 
RNFL (left column) and total (right column) thickness values at baseline (0 weeks) and after optic nerve crush (solid) compared to contralateral eyes (dashed) at ONH scans (scan 15) and superior to the ONH (scan 21). Data are represented as mean ± 95% confidence interval. Asterisks denote a statistically significant difference between control and injured eyes at different levels: *P ≤ 0.05, **P ≤ 0.005, ***P ≤ 0.0005, ****P ≤ 0.0001.
Figure 7.
 
RNFL (left column) and total (right column) thickness values at baseline (0 weeks) and after optic nerve crush (solid) compared to contralateral eyes (dashed) at ONH scans (scan 15) and superior to the ONH (scan 21). Data are represented as mean ± 95% confidence interval. Asterisks denote a statistically significant difference between control and injured eyes at different levels: *P ≤ 0.05, **P ≤ 0.005, ***P ≤ 0.0005, ****P ≤ 0.0001.
Discussion
RetOCTNet is a powerful tool for segmenting and measuring RNFL and total retinal thickness from OCT images of healthy and RGC-injured eyes in rats. It can segment the RNFL and total retinal thickness with an overall F1 score of 0.88 and 0.98, respectively. The F1 score of 0.78 for the ONC 12-week test set was the lowest across data sets, reflecting the difficulty in segmenting advanced damage near the ONH. However, our human annotators encountered similar challenges in segmenting the ONH after OHT, which likely propagated uncertainty and potentially lowered the quality of their annotations. Nevertheless, this F1 value of 0.78 in the test set for the 12-week ONC eye denotes a substantial overlap, confirming the algorithm's efficacy in capturing relevant structural information despite advanced injury stages. 
The challenge of assessing retinal thickness after injury was consistent in the uncertainty maps that highlighted the largest uncertainty was located around the ONH and highest after RGC injury. To ensure that the good F1 score, precision, and recall metrics reflect the accurate thickness measurements, we compared the ground-truth and RetOCTNet thickness measurements. We found a good overall correlation between RetOCTNet and ground-truth segmentations. The Bland–Altman plots calculated that the mean absolute offset between ground-truth and RetOCTNet thickness measurements (bias ≤1.7 µm) was less than the axial resolution of our images (resolution = 2.75 µm). This indicates that the thickness measurements agreed with ground truth and were accurate within the resolution of the OCT measurements. In addition, there were no trends in the difference measurements toward underestimating or overestimating the thickness values. We also did not find a significant difference in the final RNFL and retinal thickness measurements between RetOCTNet and ground truth. We examined segmentations where RetOCTNet did not perform well. Typically, these images had regions of low contrast (Supplementary Figs. S3 and S4) that RetOCTNet identified as mislabeled background in human annotated images or the presence of complex morphology due to the obliquely sliced blood vessels after inducing OHT (Supplementary Fig. S5). These outliers were also partially related to human annotators attempting to segment unclear regions caused by shadows, while RetOCTNet avoided segmenting these regions (Supplementary Figs. S3 and S4). This may account for large differences in thickness between human annotators and RetOCTNet. This indicates the potential need for visual inspection or postprocessing. Postprocessing can help remove areas of discontinuity or account for the ONH where the RNFL and retinal layers are no longer present. However, overall RetOCTNet was able to accurately segment the RNFL and total retina. 
In this work, we used uncertainty as a visualization tool to understand areas that were more difficult for RetOCTNet to segment. We found that there is more uncertainty when there is advanced damage, suggesting larger data sets may be needed to improve the DL approach when including advanced damage. Interestingly, expert human annotators also found it more challenging to segment RGC-injured eyes when there was advanced damage. By contrasting uncertainty between the training and test data sets, we observed similar uncertainty values across data sets and injury levels. This served as a quality control, demonstrating a lack of overfitting when the uncertainty values in the test data set were similar in magnitude to those in the training set. In future works, we plan on using uncertainty to guide the segmentations and decrease the number of required scans for training.14 In addition, most of the uncertainty was around the ONH. While this region is relevant for remodeling after RGC injury and was used for assessing the quality of RetOCTNet segmentations, it was not used for assessing RNFL or retinal thickness. Therefore, future implementations of RetOCTNet that focus on the loss or thinning of these layers may wish to use postprocessing to omit this region. This would remove the region with the most uncertainty and reduce variability in the thickness maps due to the complex geometry around the ONH (crowding of the blood vessels, thickening of the layers, and loss of retinal layers as the RGC axons exit the posterior eye). 
To further test RetOCTNet, we compared the total retinal thickness of RetOCTNet relative to human annotators on a mouse data set without RGC injury, detailed in Supplementary Section 1. These OCT scans altered the species (mouse vs. rat), scan type (annular vs. radial), scan size (1, 1.2, or 1.4 mm vs. 3 mm), and frames per B-scan (48 frames vs. 20 frames) from what RetOCTNet was originally trained on. RetOCTNet performed well on 95.7% of these scans with an F1, precision, and recall that showed excellent agreement with values of 0.97, 0.99, and 0.96, respectively. The remaining scans were poorly segmented (>25 µm difference in thickness), which suggests that human inspection is required when using RetOCTNet outside the scope of its original training. 
A potential limitation of RetOCTNet is the impact of uncertainty in the human annotations. Most deep learning approaches require human annotations or input for training and evaluation. Therefore, if an annotator has uncertainty in their segmentations, this would be propagated to the labels. However, using expert annotators and larger data sets helps minimize the potential impact of small uncertainties in the segmentations performed by humans for the training and evaluation of RetOCTNet. In future work, we plan to use uncertainty as a guide to improve segmentation accuracy and decrease the number of scans required for training. 
To ensure the ability of RetOCTNet to evaluate RNFL and retinal thickness outside the ONH, we also evaluated RetOCTNet on volume scans away from the ONH. We found RetOCTNet was able to accurately segment OCT volume scans and B-scans with or without the ONH region. This improves the use of RetOCTNet on other data sets, resolutions, and potentially for groups who want to determine RNFL and retinal thickness away from the ONH. 
A key aspect of RetOCTNet is its ability to generalize to OCT scans taken from healthy and RGC-injured eyes. Further, we developed this approach for both OHT and ONC models of RGC injury. To the best of our knowledge, there was no other deep learning–based algorithm available to do this in rats. While tuning the hyperparameters of RetOCTNet and training the network is computationally expensive, its usage is not. The algorithm can process an OCT scan within a matter of seconds. 
Despite RetOCTNet's ability to quickly analyze RNFL and total retinal thickness from OCT images, there are some limitations. One of the main challenges to address was the imbalanced data set. To avoid information leakage, scans from each eye were only present in one data set (training, testing, or validation). We also ensured that the training data set included OCT scans from each time point and condition: control eyes, OHT 4 weeks postinjury, OHT 8 weeks postinjury, and optic nerve crush. Although the overall data set did not have the same number of controls and injury types, this comprehensive distribution in the training set allowed RetOCTNet to be exposed to the various types and degrees of injury within our data set. We recognize the absence of OHT samples at the 8-week time point within the test set. Despite this limitation, the model performed well in segmenting OHT samples at the 8-week interval in the validation set, underscoring that it has even OCT scans within the training set. A similar statement can be made to the lack of ONC 12-week data in the validation set. Since the model performs adequately at this condition in the test set, that is indicative enough that the learning process was successful. Further, this different data set was composed of volumes with less frame averaging, including ONC injured eyes at 0, 4, 8, and 12 weeks, and RetOCTNet performed well at each time point. The predominance of injured eyes amplifies the algorithm's proficiency in handling complex segmentation tasks. The test and validation sets, despite their imbalance, fulfill their roles by providing a diverse range of cases, affirming the model's adaptability and robustness across varying degrees of retinal injury. 
Another limitation was that the model was exclusively trained on patches from nonoverlapping image regions. While patch-based training may introduce segmentation discontinuities, especially near the ONH, the model's high F1 scores suggest this is not a significant concern. Furthermore, as ONH thickness measurements are typically excluded in glaucoma injury assessments, any minor inconsistencies in this region do not detract from the model's overall utility. We also acknowledge that it was exclusively trained on scans from a Bioptigen OCT machine. Future work will include scans acquired with other OCT machines (e.g., Spectralis OCT Heidelberg Engineering GmbH [Heidelberg, Germany] and Phoenix Micron IV system [Bend, Oregon, USA]) to account for interinstrumentation variability and image characteristics. 
Conclusions
In conclusion, the RetOCTNet algorithm is a fully automated framework to segment the RNFL and total retinal thickness in spectral-domain OCT rodent images tested on radial and volumetric OCT images. It represents an advancement in the segmentation and measurement of RNFL and total retinal thickness from OCT images of healthy and RGC-injured rats that shortens the segmentation time needed for manually segmenting each OCT image. The algorithm demonstrated performance on par with human experts, as evidenced by the calculated F1 scores, making it a promising tool for preclinical studies of RGC injury. The RetOCTNet algorithm is publicly available for use at https://github.com/afeola2/RetOCTNet.git
Acknowledgments
The authors thank Research to Prevent Blindness, Inc., for a Challenge Grant to the Department of Ophthalmology at Emory University. 
Supported by the National Institutes of Health (NIH) NEI to AJF (R21 EY035468 & R01 EY030871) and MTP (R01 EY016435 & R01 EY033361), Department of Veterans Affairs Research Career Scientist Award RX003134 (MTP), and NIH NEI P30 core grant P30EY006360. The project that gave rise to these results also received the support of a fellowship from “La Caixa” Foundation (ID 100010434). The fellowship code is LCF/BQ/EU21/11890105. 
Data underlying the results presented in this article are not publicly available at this time but may be obtained from the authors upon reasonable request. 
While writing this manuscript, the authors used ChatGPT to improve the writing of some paragraphs. After using this tool, the authors reviewed and edited the content. The authors take full responsibility for the content of this manuscript. 
Disclosure: G. Sanchez-Rodriguez, None; L. Lou, None; M.T. Pardue, None; A.J. Feola, None 
References
Quigley HA, Broman AT. The number of people with glaucoma worldwide in 2010 and 2020. Br J Ophthalmol. 2006; 90: 262–267. [CrossRef]
Tham Y-C, Li X, Wong TY, et al. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology. 2014; 121: 2081–2090. [CrossRef]
Dietze J, Blair K, Havens SJ, Adams M. Glaucoma (Nursing). 2022. In StatPearls.
Budenz DL, Anderson DR, Varma R, et al. Determinants of normal retinal nerve fiber layer thickness measured by stratus OCT. Ophthalmology. 2007; 114: 1046–1052. [CrossRef]
Huang D, Swanson E, Lin C, et al. Optical coherence tomography. Science. 1991; 254(5035): 1178–1181. [CrossRef]
Hee MR . Optical coherence tomography of the human retina. Arch Ophthalmol. 1995; 113: 325. [CrossRef]
Bussel II, Wollstein G, Schuman JS. OCT for glaucoma diagnosis, screening and detection of glaucoma progression. Br J Ophthalmol. 2014; 98: ii15–ii19. [CrossRef]
Pekala M, Joshi N, Liu T, et al. OCT segmentation via deep learning: a review of recent work. In: ACCV 2018 Workshops: 14th Asian Conference on Computer Vision. Perth, Australia: Springer International Publishing: 2018;316–322.
Li K, Wu X, Chen D, Sonka M. Optimal surface segmentation in volumetric images—a graph-theoretic approach. IEEE Trans Pattern Anal Mach Intell. 2006; 28: 119–134.
Garvin MK, Abramoff MD, Wu X, et al. Automated 3-D intraretinal layer segmentation of macular spectral-domain optical coherence tomography images. IEEE Trans Med Imaging. 2009; 28: 1436–1447. [CrossRef]
Ran AR, Tham CC, Chan PP, et al. Deep learning in glaucoma with optical coherence tomography: a review. Eye. 2021; 35: 188–201. [CrossRef]
Pekala M, Joshi N, Liu TA, et al. Deep learning based retinal OCT segmentation. Comput Biol Med. 2019; 114: 103445. [CrossRef]
Jégou S, Drozdzal M, Vazquez D, et al. The one hundred layers tiramisu: fully convolutional densenets for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. Honolulu, Hawaii, USA: IEEE; 2017;11–19.
Sedai S, Antony B, Rai R, et al. Uncertainty guided semi-supervised segmentation of retinal layers in OCT images. In: Shen D Liu T Peters TM, et al., eds. International Conference on Medical Image Computing and Computer-Assisted Intervention. Shenzhen, China: Springer; 2019: 282–290.
Lazaridis G, Xu M, Afgeh SS, Montesano G, Garway-Heath D. Bio-inspired attentive segmentation of retinal OCT imaging. In: Fu H Garvin MK MacGillivray T Xu Y Zheng Y, eds. Ophthalmic Medical Image Analysis: 7th International Workshop, OMIA 2020, Held in Conjunction with MICCAI 2020, Proceedings 7. Lima, Peru: Springer; 2020: 1–10.
He Y, Carass A, Liu Y, et al. Structured layer surface segmentation for retina OCT using fully convolutional regression networks. Med Image Anal. 2021; 68: 101856. [CrossRef]
Viedma IA, Alonso-Caneiro D, Read SA, Collins MJ. OCT retinal and choroidal layer instance segmentation using mask r-CNN. Sensors. 2022; 22: 2016. [CrossRef]
He K, Gkioxari G, Dollár P, Girshick R. Mask r-CNN. In: Proceedings of the IEEE International Conference on Computer Vision. Venice, Italy: IEEE; 2017: 2961–2969.
Antony BJ, Kim BJ, Lang A, et al. Automated segmentation of mouse OCT volumes (ASiMOV): validation & clinical study of a light damage model. PLoS ONE. 2017; 12: 1–17. [CrossRef]
Mishra A, Wong A, Bizheva K, Clausi DA. Intra-retinal layer segmentation in optical coherence tomography images. Opt Express. 2009; 17: 23719–23728. [CrossRef]
Srinivasan PP, Heflin SJ, Izatt JA, et al. Automatic segmentation of up to ten layer boundaries in SD-OCT images of the mouse retina with and without missing layers due to pathology. Biomed Optics Express. 2014; 5: 348–365. [CrossRef]
Morales S, Colomer A, Mossi JM, et al. Retinal layer segmentation in rodent OCT images: local intensity profiles & fully convolutional neural networks. Comput Methods Programs FBiomed. 2021; 198: 105788. [CrossRef]
Feola AJ, Fu J, Allen R, et al. Menopause exacerbates visual dysfunction in experimental glaucoma. Exp Eye Res. 2019; 186: 107706. [CrossRef]
Allen RS, Douglass A, Vo H, Feola AJ. Ovariectomy worsens visual function after mild optic nerve crush in rodents. Exp Eye Res. 2021; 202: 108333. [CrossRef]
Han YX, Guo XX, Wang YX, Jonas JB, Chen X, Wang XF. Automated posterior scleral topography assessment for enhanced staphyloma visualization and quantification with improved maculopathy correlation. Transl Vis Sci Technol. 2024; 13(10): 41. [CrossRef]
de Vente C, Valmaggia P, Hoyng CB, et al. Generalizable deep learning for the detection of incomplete and complete retinal pigment epithelium and outer retinal atrophy: a Macustar report. Transl Vis Sci Technol. 2024; 13(9): 11. [CrossRef]
Benetz BAM, Shivade VS, Joseph NM, et al. Automatic determination of endothelial cell density from donor cornea endothelial cell images. Transl Vis Sci Technol. 2024; 13(8): 40. [CrossRef]
Rahman F . keras-contrib: Keras community contributions. 2019, https://github.com/keras-team/keras-contrib.
Biewald L . Experiment tracking with weights and biases. 2020, https://wandb.com. Accessed December 1, 2024.
Snoek J, Larochelle H, Adams RP. Practical Bayesian optimization of machine learning algorithms. Adv Neural Inform Process Syst. 2012; 25: 2951–2959.
McIntire M, Ratner D, Ermon S. Sparse Gaussian processes for Bayesian optimization. In: Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence. 2016: 517–526.
Mitson R . What is Bayesian hyperparameter optimization? With tutorial, https://wandb.ai/wandb_fc/articles/reports/What-Is-Bayesian-Hyperparameter-Optimization-With-Tutorial—Vmlldzo1NDQyNzcw. Accessed July 27, 2024.
Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980. 2014. Accessed December 1, 2023
Muller D, Soto-Rey I, Kramer F. Towards a guideline for evaluation metrics in medical image segmentation. BMC Res Notes. 2022; 15(1): 210. [CrossRef]
Li X, Sun X, Meng Y, Liang J, Wu F, Li J. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855. 2019. Accessed July 27, 2024.
Figure 1.
 
Example of RetOCTNet and human segmentations overlaid on a control eye, an optic nerve crush eye 12 weeks postinjury, and an ocular hypertensive eye 8 weeks postinjury.
Figure 1.
 
Example of RetOCTNet and human segmentations overlaid on a control eye, an optic nerve crush eye 12 weeks postinjury, and an ocular hypertensive eye 8 weeks postinjury.
Figure 2.
 
Example of RetOCTNet segmentation on OHT samples at 4 weeks post-OHT. Ground truth represents the manually annotated mask, and the prediction map shows the segmentation performed by RetOCTNet. Examples from the training set (two left columns) and test set (two right columns).
Figure 2.
 
Example of RetOCTNet segmentation on OHT samples at 4 weeks post-OHT. Ground truth represents the manually annotated mask, and the prediction map shows the segmentation performed by RetOCTNet. Examples from the training set (two left columns) and test set (two right columns).
Figure 3.
 
Correlation of the RNFL (left column) and total retinal (right column) thickness from RetOCTNet annotations versus ground truth (human annotations) for the training and test data set. The black dashed line is the unity line, the solid purple line represents the least squares solution to the best fit, and the hollow circles represent the individual thickness measurements.
Figure 3.
 
Correlation of the RNFL (left column) and total retinal (right column) thickness from RetOCTNet annotations versus ground truth (human annotations) for the training and test data set. The black dashed line is the unity line, the solid purple line represents the least squares solution to the best fit, and the hollow circles represent the individual thickness measurements.
Figure 4.
 
Bland–Altman plots from RNFL (left column) and total retinal (right column) thickness measurements of RetOCTNet and ground-truth annotations from the training and test data set. The purple solid lines represent the bias while the dashed lines are the 95% limits of agreement.
Figure 4.
 
Bland–Altman plots from RNFL (left column) and total retinal (right column) thickness measurements of RetOCTNet and ground-truth annotations from the training and test data set. The purple solid lines represent the bias while the dashed lines are the 95% limits of agreement.
Figure 5.
 
Uncertainty projection in the validation data set. The ONH region is the region with the largest uncertainty values.
Figure 5.
 
Uncertainty projection in the validation data set. The ONH region is the region with the largest uncertainty values.
Figure 6.
 
Samples of OCT scans and masks produced by RetOCTNet from OCT volumes at two different locations: superior to the ONH (scan: 21/29) (top) and at the ONH (scan 15/29) (bottom).
Figure 6.
 
Samples of OCT scans and masks produced by RetOCTNet from OCT volumes at two different locations: superior to the ONH (scan: 21/29) (top) and at the ONH (scan 15/29) (bottom).
Figure 7.
 
RNFL (left column) and total (right column) thickness values at baseline (0 weeks) and after optic nerve crush (solid) compared to contralateral eyes (dashed) at ONH scans (scan 15) and superior to the ONH (scan 21). Data are represented as mean ± 95% confidence interval. Asterisks denote a statistically significant difference between control and injured eyes at different levels: *P ≤ 0.05, **P ≤ 0.005, ***P ≤ 0.0005, ****P ≤ 0.0001.
Figure 7.
 
RNFL (left column) and total (right column) thickness values at baseline (0 weeks) and after optic nerve crush (solid) compared to contralateral eyes (dashed) at ONH scans (scan 15) and superior to the ONH (scan 21). Data are represented as mean ± 95% confidence interval. Asterisks denote a statistically significant difference between control and injured eyes at different levels: *P ≤ 0.05, **P ≤ 0.005, ***P ≤ 0.0005, ****P ≤ 0.0001.
Table.
 
RetOCTNet Performance Metrics Across Train, Test, and Validation Sets
Table.
 
RetOCTNet Performance Metrics Across Train, Test, and Validation Sets
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×