**5. The future of optical coherence tomography angiography in glaucoma**

Recently, artificial intelligence (AI) techniques such as machine learning (ML) and deep learning (DL) are increasingly being used in conjunction with OCTA to improve glaucoma diagnosis and track disease progression. However, further work is required to standardize the nomenclature and datasets utilized for ML and DL to maximize the impact of AI technology in OCTA.

#### **5.1 Standardization of optical coherence tomography angiography**

A major barrier that has prevented the compilation of large OCTA databases for ML and DL is the lack of standardization among OCTA instruments, imaging protocols, data analysis methods, and inconsistent nomenclature. This has made it very challenging to apply new technologies or methods to analyze data from different OCTA scanners and ophthalmology clinics. Therefore, establishing standardized protocols for imaging, data analysis, and terminology is crucial [54, 55].

#### **5.2 Machine learning using optical coherence tomography angiography imaging**

#### *5.2.1 Overview of machine learning methodology*

Broadly speaking, AI refers to computer science techniques which simulate human cognitive processes such as visual perception, speech recognition, and decisionmaking. ML is a subfield of AI that allows software systems to automatically learn and improve from experience without being explicitly programmed through the use of algorithms and statistical models that analyze and draw inferences from patterns in data. It involves training a model on a dataset and allowing the model to make predictions or decisions without human intervention [56, 57].

ML requires four sequential steps: an input of high-volume and high-quality data, extraction of features, model building, and performance evaluation [56]. The quality of the manual feature extraction from a dataset to be used as an input for ML is critical for optimal model performance. During this process, unique properties and patterns are identified which the model then uses to learn how to perform the task. These can include higher level features, such as glaucomatous eyes, or lower-level features such as image edges and shapes. The main caveat to high quality feature extraction is that it requires large data sets and many hours of precise identification of objects of interest by the researchers. One approach to reduce this burden is to use DL automated feature learning, where the model learns to extract and identify relevant features from a dataset automatically, without human intervention.

Once the features have been extracted and the model has been chosen, the next step is to train the model on a data set. During this process, the model's parameters are adjusted to minimize the error between the model's predictions and true value. The success of an ML model is usually evaluated in three categories: performance, resources required, and prediction accuracy [56]. Of these three, accuracy is perhaps the most important and difficult metric to assess because it requires an additional independent test dataset. The test dataset is given to the ML model and the predictions are compared against the true value or ground truth. If the model's performance is not satisfactory, it may be necessary to fine-tune the model by adjusting its hyperparameters or by training the model on additional data sets. Once the model has been trained and evaluated, it can be deployed to make predictions on new, unseen data.

#### *5.2.2 Using machine learning for optical coherence tomography angiography analysis*

Recently, ML image analysis methods have become an area of interest in OCTA research. Specifically, researchers aim to use AI to accurately detect pathology, precisely quantify retinal vasculature, and reliably diagnose disease [58]. In 2019, Chan et al. used features of the ONH and retina to automatically diagnose glaucoma from macular and disk images [59]. In this pilot study, the authors used the AdaBoost Classifier model, a self-adjusting classification tool. They reported excellent model accuracy of 94.3% on 109 images (57 normal, 52 glaucoma). This milestone study showed that ML could aid clinicians in glaucoma detection at an early stage.

In 2022, Kooner et al. developed an ML tool to identify the parameters which provided the most accurate diagnosis of glaucoma [60]. The authors analyzed six ML algorithms, and over 2500 ML models were optimized using random search. In this study, the XGBoost algorithm, a highly effective and scalable ML model, achieved

#### *The Role of Optical Coherence Tomography Angiography in Glaucoma DOI: http://dx.doi.org/10.5772/intechopen.110272*

the highest accuracy of 83.9%. They also explored the ML decision tree models to understand the most useful diagnostic parameters (inferior temporal VD, inferior hemisphere VD, and peripapillary RNFL thickness).

#### *5.2.3 Overview of deep learning methodology*

DL is a subset of ML which uses artificial networks composed of "neurons" or nodes layered to resemble the human brain. DL algorithms can understand complex patterns and relationships in the data by adjusting the weights and biases of the connections between the neurons in the network. While DL architectures are more powerful learners than ML algorithms, they are less customizable and interpretable [56]. Additionally, DL models are very practical since they are able to automatically extract features from raw input data resulting in increased efficiency and pattern recognition. Convolutional neural networks (CNNs), a type of DL architecture specialized for image input data, are particularly useful for extracting features for DL training. They operate in a bottom-up manner, first identifying basic features such as corners and working up towards more complex structures. Similar performance metrics are utilized for both ML and DL models.

A disadvantage of DL is the need for a large volume of clinical data during the training process. This data acquisition process can be affected by privacy concerns and time constraints. DL tools which can be used to address this need for data are known as generative adversarial networks (GANs). GANs are a type of DL algorithm that are used for generating new, synthetic data that is similar to a given dataset. GANs are composed of two neural networks: a generator network, which is responsible for generating new data, and a discriminator network, which can distinguish the synthetic and real data. The goal of GANs is to generate synthetic data that is indistinguishable from real data, addressing issues posed in conventional data acquisition processes.

#### *5.2.4 Using deep learning for optical coherence tomography angiography analysis*

In recent years, the use of CNNs for automated glaucoma diagnosis has grown in popularity over prior ML techniques. In 2022, Bowd et al. used CNNs to improve the performance of feature-based gradient boosting classifier (GBC) analysis, an ML technique that combines multiple subsets of models to create a powerful classification tool, in 405 images (130 healthy, 275 glaucoma) [61]. GBC models were separately trained on OCT and OCTA scans of the ONH, while the CNN model was trained solely on region proposal classifier (RPC), a type of DL architecture used for object detection tasks. To account for the imbalance of healthy and glaucomatous eyes, areas under the precision recall curves (AUPRC) were computed to evaluate the performance of the two models. The CNN model had an AUPRC of 0.97, compared 0.93 for the best ML model, indicating that the DL models improve on feature-based ML models for classifying healthy and glaucomatous eyes.

In 2022 Kumar et al. used GANs to create synthetic OCT circumpapillary images, evaluate them for gradeability and authenticity, and use them to train DL models [62]. The researchers created two models to generate both healthy and glaucomatous synthetic OCT images of the circumpapillary ONH. The optimal DL network trained on synthetic images (AUC = 0.97 internal test data vs. 0.90 external test data), while the DL network trained with real images performed worse (AUC = 0.96 internal test data vs. 0.84 external test data). The accuracy of DL networks trained with synthetic

images were comparable to those trained with real images, indicating their potential use for other modalities such as OCTA and similar DL applications.

DL can also be used to enhance the quality of OCT and OCTA images affected by artifacts and speckle noise (noise caused by the coherent scattering of the light waves used to acquire the image). Early studies by Yamashita et al. used modified CNN tools to enhance the scan quality of noisy ONH OCTs [63]. Recently, Omodoka et al. denoised RPC OCTA images to improve the quality of calculated RPC vessel area density and vessel length density [64].

Since DL automated feature extraction acts like a "black box" concept, studies explaining DL decision making are vital. These "explainability" studies attempt to address the interpretability of DL, which will be crucial for clinical implementation. To better understand DL decision making, Hemelings et al. conducted a study on fundus images to determine the importance of the regions outside of the ONH for DL-based glaucoma detection and vertical CDR (VCDR) [65]. The researchers trained DL models on a database of 23,930 images and compared classification accuracy. They showed that models trained on the original unaltered images (AUC = 0.94, VCDR estimation = 77%) outperformed models that were trained on images with the absence of the ONH (AUC =0.80, VCDR estimation = 37%). Thus, these results provide evidence that DL models that use areas beyond the ONH, such as VCDR, are superior in classifying glaucomatous eyes.

## **6. Conclusion**

Since the introduction of OCTA concept, the technology has increased the understanding of both the structural and vascular damage seen in patients with glaucoma. Not only is OCTA a safe, non-invasive, and quick test it provides the same structural information as OCT such as retinal and RNFL thinning, but it can also visualize relevant vascular parameters such as ONH VD and macular VD that typically undergo glaucomatous damage. However, significant limitations of OCTA such as a high prevalence of artifacts and lack of standardization across different machines currently do exist and must be accounted for while using this technology. Despite these limitations, the advantages of being able to observe both the structural and vascular damage caused by glaucoma show why OCTA is currently being adopted by ophthalmologists. In addition, the progress made in incorporating ML and DL techniques with OCTA will aid both in the diagnosis and progression in glaucoma.
