Machine Learning and Deep Learning Techniques for Optic Disc and Cup Segmentation – A Review

Introduction

Glaucoma is an eye disease that damages the optic nerve and causes visual impairment (7.7 million).¹ According to the World Health Organization (WHO), an estimated 2.2 billion people worldwide have a vision impairment, including 1 billion classified as moderate to severe. Various reports have been published on glaucoma prevalence worldwide (Table 1).

Table 1 Summary of Glaucoma Prevalence Studies Around the World

Glaucoma is known as the “silent theft” of sight because the associated vision loss results from increasing ocular pressure on the optic nerve, which is often asymptomatic. Thus, early diagnosis of glaucoma is essential to prevent irreversible vision loss. One major factor in the early detection of glaucoma is identifying optic nerve damage, which can be detected by using a fundus camera developed for this purpose.¹⁴ Fundus imaging is a non-invasive procedure that relies on the monofocal indirect ophthalmoscopy principle.

Computer vision problems have significantly benefitted from recent advancements in deep learning methods. Image segmentation is a fundamental component of computer vision and visual-understanding problems. It can be defined as the sub-division of an image or video based on a distinct visual region with semantic meaning. Over the past few years, deep learning has achieved greater accuracy rates than traditional image processing techniques and has performed well by many popular segmentation benchmarks. This has led to a paradigm shift in image segmentation.^15,16

Therefore, the paradigm shift towards deep learning in image segmentation has affected biomedical imaging applications. Currently, deep learning has become an efficient and effective methodology for a wide variety of biomedical image segmentation tasks in cardiac imaging,¹⁷ neuroimaging,^18,19 and abdominal radiology.²⁰

To date, segmentation with deep learning has been used to detect, localize, and diagnose clinical problems such as breast cancer,²¹ lung cancer,²² and Alzheimer’s Disease.²³ Deep learning segmentation has been successfully used with most major imaging modalities, including magnetic resonance imaging (MRI),²⁴ computed tomography (CT),²⁵ and ultrasound.²⁶

The use of computer-aided diagnosis (CAD) systems in ophthalmology to segment the required features has expanded over the last two decades, particularly using fundus images or optical coherence tomography (OCT) for glaucoma and retinal diseases such as diabetic retinopathy. This expanding interest has led scientists to investigate machine learning and deep learning techniques for CAD systems in recent years, which has generated a large volume of published methodologies, methods, and datasets. Thus, review studies can play a crucial role in advancing CAD systems by providing specialists with a summary of the current landscape and pathways for future investigation. Previous review studies have examined artificial intelligence (AI) and its applications in ophthalmology, both in general and in relation to specific conditions, including glaucoma.^27–30

This review is structured as follows: a comprehensive literature review is performed in the following section. After that, the material and methods section is introduced. In the fourth section, the results are presented. Finally, the fifth and sixth sections constitute the discussion and conclusion.

Literature Review

It is noteworthy that reviews describing the general use of AI in ophthalmology tend to target a broad audience, leading to limited technical conclusions with implications for specific ocular conditions. Ting et al³⁰ presented a thorough review of deep learning techniques for ophthalmology, highlighting technical and clinical domains. They provided, in detail, conventional deep learning models, training, datasets, reference standards (ground truths), and outcome measures. Furthermore, this review discussed many ocular conditions such as diabetic retinopathy, retrolental fibroplasia (RLF), age-related macular degeneration (AMD), and glaucoma. Sengupta et al³¹ discussed and compared the most important deep learning techniques with applications in ophthalmology, specifically focusing on glaucoma.

Owing to the increasing prevalence of glaucoma, many AI methods, image processing techniques, and glaucoma-related datasets have been published in recent years, along with several glaucoma-specific reviews.^32–45 Mursch-Edlmayr et al³⁵ presented a clinically focused review exploring AI-based strategies for detecting and monitoring glaucoma. This review examined various glaucoma testing modalities, such as optical coherence tomography (OCT), visual field (VF) testing, and fundus photos, with minimal technical discussion. They concluded that AI-based algorithms for analyzing fundus images have high translational potential due to the accessibility and simplicity of fundus photography. Future integration of AI-based analysis of fundus images may help address the current limitations of glaucoma management around the world.

Several technical review studies have also been published examining machine learning approaches and network architectures employed for glaucoma assessment and management. Hagiwara et al³³ reviewed CAD systems for glaucoma diagnosis based on conventional machine learning. They examined the entire CAD procedure, including pre-processing techniques, segmentation methods, feature extraction strategies, feature selection approaches, and classification methods. They concluded that integrating deep learning methods simplifies these systems and improves their reliability. Thakur and Juneja³² compared recent segmentation and classification approaches for glaucoma diagnosis from fundus images; they also discussed the limitations of these methods and possible avenues to increase their efficiency. Barros et al³⁴ reviewed machine learning approaches for diagnosing and detecting glaucoma. Their review examined supervised methods for glaucoma prediction from fundus images and categorized them into deep learning and non-deep learning approaches. Eswari and Karkuzhali’s⁴⁶ summarised the advantages and disadvantages of many segmentation and classification techniques.

This study examines and critiques the deep learning methods that have been explicitly designed for optic cup and disc segmentation to diagnose glaucoma from fundus images. This review will be considering the methods section covering the most public fundus images datasets, the evaluation metrics used to evaluate the developed deep learning segmentation approaches, and finally, the most developed machine learning and deep learning architectures for optic disc and optic cup segmentation.

Materials and Methods

Research Criteria

The required information was retrieved for glaucoma (optic disc and optic cup boundaries) machine learning and deep learning segmentation approaches. The image datasets presented in this review were obtained from various websites, conference papers, and journal peer review studies. Therefore, this review was not required to obtain ethical approval to be conducted.

Study Selection

Standard academic web search engines were used, such as Google Scholar, PubMed, and Web of Science. Glaucoma, optic disc, and cup segmentation were placed after or before the following keywords or phrases: deep learning, machine learning, artificial intelligence, telemedicine, fundus datasets, and fundus databases. The search was conducted from October 2020 to March 2021.

Results

Datasets

Multiple fundus retinal image datasets were developed for glaucoma detection. Several datasets were also developed for different retinal diseases, such as diabetic retinopathy, which this review will not consider. These images may not meet crucial criteria, such as sufficient image quality, which can subsequently limit the efficacy of machine learning analyses. In addition, these image sets may be insufficient for the development of machine learning processes for certain diseases that focus on specific areas of the retina, as is the case with the optic nerve head in glaucoma. Moreover, some datasets, such as RIM-ONE and Drishti-GS, only have a small number of images, hindering the development and testing of machine learning algorithms. Therefore, efforts have been made to develop task-oriented datasets to overcome the limitations of existing datasets and identify the images required to train algorithms to detect glaucoma automatically. This study reviews the available retinal images datasets currently used for automated glaucoma detection (Table 2).

Table 2 Current Available Public Datasets with Labelling and Manual Annotation for Glaucoma Fundus Images

Evaluation Metrics

Evaluating segmentation maps for optic disc and cup segmentation is not a straightforward task. In the literature, various evaluation metrics have been used to assess the performances of machine learning and deep learning models. The optic cup and disc are objects in segmentation tasks where both the ground truth and segmentation encompass both foreground and background partitions. The evaluation of a segmentation algorithm can focus on assessing how well the pixels of the ground truth are detected; therefore, most researchers use a pixel-wise approach to assess the efficacy of their models. To further assess the quality of the detected object, some researchers have used the boundary assessment approach.

In this section, we will discuss the evaluation metrics used in the literature. First, we need to define the relevant notation. Given a segmentation method m and a ground truth gt, denote P_m and N_m as the positive and negative pixels of an image, respectively, such that I_m = P_m N_m is the resulting object detected. Similarly, the ground truth is represented as I__gt = P_gt N_gt, where P_gt and N_gt are the positive and negative pixels, respectively. To achieve optimal detection, a model should aim for P_m= P_gt. For a pixel-based object detection measurement, there are four fundamental indicators as follows:

• True positive (TP) results represent the pixels that are labelled and predicted as an object, such that TP = P_m ∩ P_gt.
• False negative (FN) results represent the pixels that are labelled as an object but predicted as non-object, such that FN = Nm ∩ P_gt.
• False positive (FP) results represent the pixels that are not labelled as an object but are predicted an object, such that FP = P_m ∩ N_gt.
• True negative (TN) results represent the pixels that are both labelled and predicted as non-object, such that TN = N_m ∩ N_gt.

Pixel Accuracy

Pixel accuracy⁵⁶ reflects the ratio of successful object detection using the total number of true positives and true negatives to report the total number of correct predictions. This metric is a standard and simple evaluation technique that was initially developed for classification. However, pixel accuracy is used similarly for segmentation evaluation by reporting the percentage of pixels in the image that are correctly classified. The mathematical expression for pixel accuracy is as follows:

(1)

Generally, pixel accuracy is a misleading metric when the object representation within the image is small because the measure will be biased in reporting how well the negative case is identified. This bias is very relevant in evaluating the accuracy of optic cup detection in a fundus image, because this structure is typically a small portion of the fundus image.

Precision and Recall

Two of the most popular evaluation metrics for both classification and segmentation are precision and recall.⁵⁷ Precision, also known as specificity, evaluates how accurate the model is predicting in positive pixels; whereas recall, also known as sensitivity, measures the percentage of correctly identified true positives TP. Precision and recall can be expressed mathematically from the pixel-wise perspective as follows:

(2)

When the cost of an FP is very high and the cost of an FN is low, precision is the recommended evaluation metric. However, when the cost of an FN is high, recall is the more important metric. Generally, precision and recall are reported together in evaluating optic cup and disc segmentation.

Dice (F-Measure)

The trade-off between precision and recall can be estimated using the F-measure, commonly known as DICE, which can effectively combine precision and recall into one formula⁵⁸ and has since become a popular evaluation technique for many segmentation problems. In fact, DICE has been used as the official evaluation metric for the widely known Retinal Fundus Glaucoma Challenge Edition (REFUGE) challenge.⁴⁸ F-measure can be expressed as

(4)

Jaccard Index

The Jaccard index,⁵⁹ also known as the intersection of union (IoU), is one of the most common evaluation metrics in image segmentation. This coefficient is defined as the size of the intersection divided by the size of the union of the sample sets:

(5)

Both the Jaccard index and DICE provide similar results; in other words, if one metric says that Model A offers better segmentation than Model B, the other metric will produce the same result. However, there are some minor differences between the metrics, such as the penalisation of single instances of misclassifications. Because this study focuses on the general concepts of these metrics, we refer the reader to the following studies for more details.^60,61

Boundary Distance Localisation

One way to evaluate segmentation methods is to consider a boundary approach instead of a pixel-based approach. A pixel-based approach measures how well the pixels of the ground truth are detected while a boundary-based approach focuses on how accurately the boundaries are represented.⁶⁰ The boundary approach can better reflect the quality of object detection than the pixel-based approach in some object detection problems, such as optic disc segmentation. A common boundary-based evaluation metric is the mean boundary distance in pixels between the predicted segmentation results and the ground truth. This measure is known as boundary distance localisation (BDL). BDL can be defined mathematically as

(6)

where and are the distance from the center of the cup or disc to the predicted and ground truth boundaries, respectively. This evaluation method was used by Jiang et al⁶² and produced high marginal variability for disc segmentation when compared to the traditional pixel-based metric, such as F-measure.

An objective assessment of different proposed methods for OC and OD segmentation is presented in Table 3 using mean dice, sensitivity, specificity, and Jaccard on the REFUGE test set. In the table, the comparison shows that the state-of-the-art graph convolutional network based (GCN)⁶³ achieved a mean dice coefficient of 95.58% and 97.76% for OD and OC, respectively.

Table 3 Comparison of Optic Cup (OC) and Optic Disc (OD) Segmentation Methods on REFUGE Test Set Using Different Metrics

Machine Learning and Deep Learning Architectures

The abundance of newly generated datasets coupled with growing computing power has led to diverse deep learning approaches to image analysis. Image segmentation has benefited from this advancement, revolutionizing traditional image processing techniques.⁶⁵ More deep learning approaches have begun to appear specifically for segmentation because the nature of segmentation tasks have expanded to applications for autonomous vehicles,⁶⁶ augmented reality, video surveillance, and medical imaging. This has improved the advancement of deep learning.

As an important task in medical image analysis, fundus image segmentation has become a significant contributor to the evolution of deep learning methods. In this section, we will review different deep learning architectures that have been employed for optic cup and disc segmentation (Table 4). We divide these methodologies into four major categories: convolutional neural network-related approaches, U-Net-related methods, generative adversarial network (GAN) approaches, and other deep learning-related approaches.

Table 4 Deep Learning and Machine Learning Methods for Optic Cup and Disc Segmentation

A Convolutional Neural Network (CNN)

CNN is a type of feed-forward artificial neural network (ANN). CNN’s comprise stacked layers of convolution and pooling operations, followed by batch normalization layers to speed up learning and stabilize the input of subsequent layers as the network grows deeper. In the final layers, weight vectors are concatenated and passed into one or more fully connected layers for classification output. The convolution layers slide over the input with a fixed step and window size to control the moving dot product of the data. The flattened output is then passed into a differentiable nonlinear activation function, typically a rectified linear unit (ReLU). As a result of its success in computer vision tasks over state-of-the-art conventional machine learning algorithms, such as support vector machines, many researchers have investigated the ability of CNNs to segment the optic cup and disc regions,^93–95 and retinal blood vessels⁹⁶ in fundus images. Unlike image-level classification or bounding box-level prediction tasks, the feature maps in semantic segmentation are resized to the original image space. Each pixel is given a probability of being assigned to a given semantic label.

Tan et al⁹⁴ used a CNN with two convolutions followed by a max-pooling layer that connects to a fully connected layer with 100 neurons and a final layer of 4 neurons as an output to classify each pixel of a fundus image as either background, optic cup, fovea, or blood vessel. Figure 2 shows the authors’ proposed CNN architecture (Figure 2). Sreng et al⁹³ employed DeepLab-v3+,⁹⁷ which comprises an encoder and decoder module, for optic disc segmentation. The network design uses spatial pyramid pooling (Figure 1), accounting for different optic disc sizes across fundus images with varying scales. The authors built and evaluated their optic disc semantic segmentation models with 2787 retinal images from five different publicly available datasets.

Figure 1 Tan et al proposed CNN architecture. Notes: Reprinted from: Tan JH, Acharya UR, Bhandary SV, Chua KC, Sivaprasad S. Segmentation of optic disc, fovea and retinal vasculature using a single convolutional neural network. J Comput Sci. 2017;20:70–79.94 Copyright 2017, with permission from Elsevier.

Figure 2 Spatial pyramid pooling layer: pooling features extracted using different window sizes on the feature maps. Notes: © 2015 IEEE. Reprinted, with permission, from: He K, Zhang X, Ren S, Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell. 2015;37(9):1904–1916.98.

Sun et al⁸² reframed optic disc segmentation as an object detection problem because the shape of the optic disc is a non-rotated ellipse. They used Faster R-CNN,⁹⁹ with VGG-16 as a backbone for the R-CNN.

Lu et al¹⁰⁰ explored weakly supervised learning. This approach aimed to stochastically cluster every pixel through pairwise pixel information into the background and foreground for optic disc segmentation using two fully connected layers. The network was simultaneously trained on three different tasks: 1) optic disc segmentation, 2) glaucoma classification, and 3) evidence map (heatmap for affected regions) prediction.

Shankaranarayana et al¹⁰¹ proposed a pipeline for an end-to-end, fully connected convolutional encoder-decoder network with two paths for the encoder component of the network. This pipeline took two modalities as input: an RGB fundus image and a depth map. The authors used a pre-trained network to predict the depth map for monocular retinal depth estimation. Subsequently, the original image and the predicted depth map were fed into another encoder-decoder network for optic cup and disc segmentation.

U-Net Based Approaches

Unlike other machine learning and deep learning methodologies, U-Net, a fully convolutional network (FCN) sub-type, was explicitly developed for biomedical imaging segmentation.¹⁰² The architecture and training strategy of a U-Net-based approach promotes the efficient use of annotated data for more reliable prediction. The architectural design is a compression path that consists of consecutive convolutional layers and a max-pooling layer to extract features while limiting the feature map size (Figure 3: U-Net Architecture). Skip connections, an alternative path for backpropagation, were used to share localization information and expand layers.

Figure 3 U-net architecture (example for 32×32 pixels in the lowest resolution). Each blue box corresponds to a multi-channel feature map. The number of channels is denoted on top of the box. The x-y-size is provided at the lower left edge of the box. White boxes represent copied feature maps. The arrows denote the different operation. Notes: Reprinted by permission from Springer Nature from: Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells W, Frangi A (eds). Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham; 2015:234–241.102 Copyright © Springer International Publishing Switzerland 2015. Available from: https://link.springer.com/book/10.1007/978-3-319-24574-4.

This unique architecture captured attention after winning multiple segmentation challenges in 2015 and has since been widely used, especially for medical applications. Most neural network-based fundus segmentation methods for glaucoma have relied on U-Net or U-Net alternatives. Sevastopolsky⁸¹ modified U-Net and used the Jaccard Index and DICE score to compare their methods with two different CNN-based approaches.^76,103

Chakravarty and Sivaswamy⁷⁰ proposed a glaucoma assessment framework that jointly segmented and classified fundus images. They proposed a multi-task CNN architecture wherein U-Net is used for segmentation, and the output and latent layers are used together to predict glaucoma. Al-Bander et al⁷² proposed DenseNet, which simultaneously segments the optic cup and disc. DenseNet resembles the traditional U-Net architecture with some differences, such as a Dense Block, transition Down Block, and transition Up Block. They presented a comprehensive comparison using multiple datasets and evaluation metrics to compare their method with deep learning-based and traditional image processing techniques.

Fu et al⁷³ presented the Disc Aware Ensemble Network (DENet) for glaucoma screening that incorporates a global fundus image and an image that focuses on the optic disc region. However, as the DENet was used for glaucoma screening, the U-Net network was used for segmenting the disc to influence the screening decision. As such, their method was evaluated only for glaucoma screening rather than segmentation performance. Fu et al⁷⁸ also introduced a novel joint optic cup and disc segmentation approach based on a polar transformation of the fundus image. To this end, they proposed M-Net, which is a multi-scale U-Net with a side-output layer that produces a probability map.

Sevastopolsky et al⁸⁷ proposed a U-Net-based, special cascade network that incorporated the concept of iterative refinement. The model was designed as stacked multiple blocks to achieve better recognition where each block was a basic U-Net. They compared their model with one U-Net-based model⁸¹ and two CNN-based approaches.^76,103

Wang et al⁶⁸ introduced a coarse-to-fine deep learning model based on U-Net designed specifically to segment the optic disc region. The coarse-to-fine strategy splits the segmentation task into multiple stages: one for extracted vessels (vessel density map) and the other for local disc patches to obtain segmentation results. The authors compared the performance of their disc segmentation method against several deep learning and image processing methods on various datasets. Although this method is novel from a prepossessing perspective, it did not consistently improve across all datasets and did not include optic cup segmentation.

Gu et al⁷¹ proposed a context encoder network (CE-NET) for medical image segmentation. CE-NET modified the U-Net architecture to account for the consecutive pooling and striated convolution operation, leading to spatial information loss. They replaced the U-Net encoder block with a ResNet-34 pre-trained block and added a context extractor that consists of a dense atrous convolution block and a residual multi-kernel pooling block. This method was proposed for general medical image segmentation and was evaluated for optic disc segmentation. The model was evaluated using the standard pixel-based mean and standard deviation.

Finally, Yu et al⁸⁵ proposed a robust optic cup and disc segmentation method by modifying the U-Net architecture. To this end, they adopted the ResNet-34 pre-trained block for encoding and kept the original U-Net block for decoding. Krishna Adithya et al¹¹⁴ developed “EffUnet”, a two-phase network to detect glaucoma by first segmenting cup and disc using efficient Unet and then predicting glaucoma using a generative adversarial network.

Generative Adversarial Network Approaches

Deep learning architecture focused on decision-making features such as classification, regression, or segmentation in the past. However, generative networks such as variational autoencoders¹⁰⁴ and adversarial networks introduced a creative component to neural networks. A generative adversarial network (GAN) framework was introduced by Goodfellow et al¹⁰⁵ in 2014. GAN attracted attention owing to the simplicity of the network structure (Figure 4) and its robust generative performance. GAN essentially consisted of two main components: a generator that captured data distribution and a discriminator that estimated the probability of a certain feature. One primary application of GAN was an image-to-image translation, which is defined as translating one possible representation of a scene to another. This concept allowed GAN to be used for segmentation by decoding the input image into the required label. Many researchers have used this concept to solve major segmentation problems, such as semantic segmentation¹⁰⁶ or specific medical image segmentation problems.¹⁰⁷ Accordingly, many researchers have used this approach to segment optic cups and discs to improve the performance of current methods.

Figure 4 GAN general architecture consists of Generator (G) which output a synthetic sample given a noise variable input and a Discriminator (D) which estimate the probability of a given sample coming from real dataset. Both components are built based on neural network. Notes: © 2018 IEEE. Reprinted, with permission, from: Creswell A, White T, Dumoulin V, Arulkumaran K, Sengupta B, Bharath AA. Generative adversarial networks: An overview. IEEE Signal Process Mag. 2018;35(1):53–65.108.

Singh et al⁸⁴ adopted a conditional GAN¹⁰⁹ with a generative deep learning network that learned invariant features to segment the optic disc. A major component of the network, according to the authors, improved the segmentation performance in the skip connection to the encoder-decoder within the generator network.

Liu et al⁷⁹ introduced a semi-supervised segmentation model based on a conditional GAN. Their model consisted of a segmentation net and a generator that increased the training set while the discriminator identified fake images to ensure compatible training. This method demonstrated the advantages of semi-supervised methods by utilizing unlabelled data, which could help mitigate the lack of labeled data. The segmentation performance model was compared to traditional U-Net, M-Net,⁷⁸ and other architectures. Wang et al⁸³ used a patch-based output space adversarial learning framework that jointly segmented the optic cup and disc. Their model exploited unsupervised domain adaptation to address the domain shift across datasets that usually reduced the performance of the segmentation model when it was trained on a specific dataset and tested on a different dataset. They also introduced a morphology-aware segmentation tool that produced better segmentation loss.

Son et al⁸⁹ introduced another GAN-based framework for retinal vessel and optic disc segmentation alternative to CNN-based methods. Their model showed significant performance improvement for retinal vessel segmentation and no improvement for optic disc segmentation. Jing et al⁶² proposed GL-NET, a hybrid deep-CNN and GAN model for segmenting the optic cup and disc. The generator was a full CNN that included an encoder-decoder to extract features using the VGG16 network.¹¹⁰ They added skip connections within the generator to promote low- and high-level feature information. To better address the variability between models, they adopted the boundary approach BDL and pixel-based evaluation metrics to compare the model’s performance.

Building a unified model that generalizes across image types was challenging due to the large variety of fundus cameras. Wang et al¹¹¹ addressed this issue by creating a domain-invariant model based on a GAN. Their model adopted domain adaptation methods built on source domain data and utilized a small target dataset to improve the performance in the target domain. The model used the DICE coefficient to compare the performance with other significant studies. Bisneto et al⁷⁵ proposed a full glaucoma detection system that utilized GAN for image segmentation. They built an entire system to detect glaucoma, their segmentation method focused on the optic disc only, and performance was evaluated accordingly.

Other Deep Learning-Related Architectures

The three approaches described above are the most common in machine learning and deep learning for glaucoma-related image segmentation. However, some researchers have investigated different strategies and architectures and have achieved exciting results. Sedai et al⁸⁸ proposed a cascaded shape regression network that learned the final shape for segmentation. They employed a boosted regression tree and proposed a data augmentation approach to improve segmentation performance. Their evaluation used the DRISHTI dataset. Zhang et al⁷⁴ presented embedded edge-attention representations to guide a segmentation network called ET-Net. Their model followed an encoder-decoder with two additional components for guidance and aggregation. The model was tested on multiple segmentation tasks, including optic disc and cup segmentation.

Ding et al⁷⁷ introduced hierarchical attention networks for medical image segmentation. The approach combined encoder-decoder networks with CNNs to extract feature maps. Three blocks were added, specifically a dense similarity block, an attention propagation block, and an information aggregation block. The model was tested on the REFUGE dataset. Tabassum et al⁶⁷ introduced an encoder-decoder network (CDED-Net) for optic cup and disc segmentation. This architecture was slightly different from the typical U-Net because there was no bottleneck layer, and fewer convolutional layers were employed.

Discussion

This review outlines the abundant deep-learning-based optic cup and disc segmentation models from fundus images. This technique has enormous potential for clinical application because it is a robust method that overcomes many of the challenges associated with traditional image processing techniques.³² Researchers widely use three main network architectures for accomplishing the optic cup and disc segmentation task; importantly, each network architecture has its advantages and disadvantages. The traditional CNN offers excellent optic cup and disc segmentation; however, it typically requires many accurately labeled images to reach its full potential. The U-Net architecture is the current state-of-the-art architecture owing to its training simplicity and data efficiency. However, its multi-scale skip connection tends to use unnecessary information, and low-level encoder features are insufficient, leading to poorer performance. The GAN method uses a creative image-image translation component to achieve accurate optic cup and disc segmentation; however, current GAN methods are deterministic, making them difficult to generalize. Finally, further investigation of generative stochastic models is needed to study the randomness effect because the natural world is stochastic.

There are some common challenges with methods for optic disc and cup segmentation from fundus images that need to be addressed to translate these tools to clinical settings:

● Reliable and broadly applicable deep learning models in computer vision commonly rely on a large set of training and testing data. However, as observed in the reviewed models, these publicly available datasets are often relatively small. Accordingly, accumulating a large set of data labeled by specialists is the most critical challenge in building a reliable and generalizable model. Although researchers have attempted to account for this issue using strategies such as transfer learning, the need for a large volume of high-quality data for large-scale evaluation of segmentation methods remains.

● Current fundus image annotation of the optic cup and disc is subjectively performed by ophthalmologists, inducing the inter– and intra-observer variability. Almazroa et al¹¹² studied this phenomenon in-depth. They concluded that the variability in annotation results from the unclear optic disc and cup boundaries are human-related factors such as examiner fatigue or lack of concentration, or image-related factors such as low quality, hazy, unfocused images, or display devices that are too dim or too bright. This variability poses a challenge for scientists building a segmentation model using a non-standardized labeling mechanism. To account for this problem, an effective segmentation model must address these differences while training and learning unambiguous weights to perform better than human identification ultimately.

● There are a variety of fundus imaging modalities that acquire the images of the eye’s posterior pole with different resolutions, angles, and degrees of the posterior segment. Furthermore, some imaging modalities require dilated pupils, whereas others require non-dilated pupils. This creates a domain shift problem, a well-known and well-defined problem in machine learning. Most of the segmentation models reviewed in this study utilize training and test data with the same image features, ignoring the significant real-life challenge posed by inter-model variability. Wang and Deng¹¹³ discussed various techniques to address this problem, essential for building a reliable automated diagnostic tool.

● Evaluating a segmentation model is not as straightforward as evaluating classification. It is difficult to compare and identify the best architecture for optic cup and disc segmentation. This challenge created variabilities among researchers who used different metrics (reviewed in section Evaluation Metrics) to assess their work. A unified evaluation criterion for segmentation models is critical to compare different studies effectively. Such standardized evaluation criteria should include an evaluation metric that considers optic cups, disc vulnerabilities, and difficulties and uses a common evaluation methodology. To this end, we suggest that boundary approaches may be more valuable in assessing the quality of detected objects than pixel-based methods.

Devising an efficient but low-performing degradation model for OC and OD segmentation (by reducing the model complexity) is of high importance^114,115 for the enhancement of both training and inference times. Additionally, developing an explainable model is another topic that many believe is crucial for Glaucoma AI-based detection tools to be approved and accepted in the clinic.^28,116

The future direction for better optic cup and disc segmentation should account for the above cases. A large set of annotated data should be published to create and evaluate better models. Because the annotation is not a straightforward task, public data must consider inter and intra-observer variability by providing multi annotations per image. More advanced machine learning methodologies must be experimented with to overcome image discrepancies to achieve a fully robust model. Lastly, a unified evaluation metric must be adopted for upcoming models. One pixel-based approach such as F-score and a boundary approach such as boundary distance localization are recommended.

Clinical Value of Automated Optic Disc and Optic Cup Segmentation

The current advancement in artificial intelligence within healthcare focuses on enhancing various aspects such as improving accessibility, increasing speed, reducing cost, early diagnosis, and enhancing human abilities. The Singapore Integrated Diabetic Retinopathy program (SiDRP) is a telemedicine-based screening program for diabetic retinopathy.¹¹⁷ The SiDRP started to integrate AI algorithms into their workflow with the ultimate goal of fully automating the screening process, increasing efficiency, and reducing the cost for the whole program.¹¹⁸

One proven application for AI within healthcare is clinical decision support. Many clinical diagnoses and screening, especially medical imaging, are time-consuming and labour-intensive. Thus, AI offers an appealing solution to expedite decision-making and improve efficiency. Various companies began to offer an AI-based clinical support system for chest x-ray screening, brain MRI, breast mammography, and diabetic retinopathy.

Glaucoma diagnosis is a complex problem requiring well-trained ophthalmologists or optometrists. Unfortunately, there is a limited number of trained specialists worldwide to examine all suspected cases. A glaucoma decision support system could offer an alternative route to this problem. Such a system could expedite the specialist’s decision-making and may eventually evolve as an automated screening system. The Cup and disc segmentation system is a major milestone and fundamental to reaching the ultimate destination. Once a system is mature enough, it will revolutionize the clinical practice to screen and diagnose glaucoma.

As been mentioned previously, cup and disc segmentation undergo an inter-observer variability. Such disagreements are related to various factors; some are human-related factors. As with the nature of humans, examiners might experience fatigue (marking at the last hour) and lack of concentration (marking in a rush or during busy hours). It has been recommended that a second opinion could improve the overall segmentation.¹¹² However, a second opinion might not be available all the time due to various reasons. Therefore, an AI-based segmentation system could be that second opinion once mature.

Conclusion

Effective optic cup and disc segmentation on fundus images are critical for accurate automated diagnosis and prediction of glaucoma. Machine learning and deep learning models have shown promising results for various segmentation tasks, both for fundus image analysis and other applications. This study presents a comprehensive review of the current deep learning and machine learning models explicitly designed for optic cup and disc segmentation on fundus images. We also reviewed the available retinal image datasets and common evaluation metrics to provide a broad understanding of how automated image segmentation tools are designed, tested, and evaluated. While three network architectures have been commonly used for optic cup and disc segmentation, the U-net architecture and the GAN model have demonstrated robust results. They may have the potential to be tested in clinical settings shortly. However, many challenges still need to be addressed to create robust image segmentation models that can be applied in clinical settings or large-scale diagnosis campaigns.