KEY CONCEPTS
- •
Machine learning models have been used effectively in the diagnosis of keratoconus.
- •
Machine learning models such as neural networks, decision trees, and support vector machines have shown excellent sensitivity and specificity in diagnosis.
- •
Data from topographers, Scheimpflug imaging analysis, and anterior segment optical coherence tomography devices can adequately feed machine learning models, producing excellent results in the diagnosis of keratoconus.
Introduction
The estimated prevalence of keratoconus (KC) has increased 5- to 10-fold over previously reported rates in population studies, and the increase is considered to be the result of a combination of earlier and more advanced tomography detection and comprehensive data collection. An early diagnosis of KC is crucial because of its significant prevalence and the ability of early intervention to obviate the need for corneal transplantation. This is particularly important, as KC and ectatic diseases represent the most common indications for transplant in pediatric patients.
Artificial intelligence (AI) is one of the major fields of computer science research. In particular, its subfield, machine learning (ML), is being applied in several industries ranging from product recommendation in e-commerce to various uses in medicine. ML has been used in ophthalmology to diagnose eye conditions such as glaucoma and diabetic retinopathy. At a very basic level, the goal is to build a concise distribution model of class labels in terms of predictive features. The resulting classifier can then be used to assign class labels in test cases where the values of the predictor features are known but that of the class label is unknown. The defining characteristic of ML algorithms is the improvement in prediction quality with increasing experience; therefore, the more data we provide, the better the resultant prediction model.
ML systems can be categorized into two types: supervised and unsupervised.
Supervised learning involves training a model with previously labeled training data, tuning the input weights to improve the accuracy of its predictions until they are optimized, and then mapping the test datasets as corresponding outputs. It may expedite the classification process and helps distinguish various clinical outcomes. In unsupervised learning , we train a model with unlabeled data (without a human-labeled process). There is no actual instructor or teacher in unsupervised learning, and the algorithm must learn to understand the data without any guide. This type of learning incorporates models that describe hidden structures that are usually invisible to humans and can lead to new discoveries.
In the unsupervised learning subtype, the most common method is deep learning (DL), also referred to as deep neural networks, which involves using an unsupervised method and multiple layers between the input and output, avoiding manual selection and segmenting the areas of study, speeding up the processes.
The most important difference between DL and traditional ML is their performance while handling increasing amounts of data. DL algorithms do not perform well with small amounts of data, because they require a large amount of data for adequate analysis. DL is applied in various AI fields such as speech recognition, image recognition, natural language processing (NLP), robot navigation systems, and self-driving cars. Fig. 13.1 presents a schematic visualization of AI.
Models of Artificial Intelligence in Keratoconus
Several AI models have been used in the diagnosis of KC and its milder forms. The ML models most frequently used to discriminate between normal and KC eyes are neural networks, decision trees, and the support vector machine (SVM) model, followed by random forests and linear discriminant analysis. Table 13.1 shows the ML models used in different studies using AI for KC diagnosis. Studies have reported up to 100% accuracy, sensitivity, and specificity using neural networks, random forests, and Bayesian networks.
| Author, Year | Device Used for Keratoconus Diagnosis | Machine Learning Models Used | ||
|---|---|---|---|---|
| Kovács et al., 2016 | Pentacam HR, OCULUS Optikgeräte GmbH | Neural network | ||
| Twa et al., 2005 | Keratron corneal topographer (v3.49; Optikon 2000, Rome, Italy) | Decision tree | ||
| Chastang et al., 2000 | EyeSys System 2000 | Decision tree | ||
| Maeda et al., 1994 | Computer-assisted videokeratoscope (TMS-1, Computed Anatomy, New York, NY) | Decision tree, linear discriminant | ||
| Karimi et al., 2018 | Corvis ST (software version 1.00r30), corneal topography (Allegro Topolyzer; WaveLight AG, Germany), corneal tomography (Pentacam) | Neural network | ||
| Chandapura et al., 2019 | OCT (RTVue, Optovue Inc., Irvine) and Scheimpflug (Pentacam, OCULUS Optikgeräte GmbH, Wetzlar, Germany) | Random forest | ||
| Maeda et al., 1995 | Videokeratoscope, points on rings 2, 3, and 4 from videokeratographs (TMS-1, Computed Anatomy Inc., New York, NY) | Decision tree, linear discriminant | ||
| Issarti et al., 2019 | Pentacam HR, OCULUS Optikgeräte GmbH | Neural network | ||
| Smolek and Klyce, 1997 | TMS-2 for Windows videokeratography system (software version W1.2) | Neural network | ||
| Smadja et al., 2013 | GALILEI system (software version 5.2.1) | Decision tree | ||
| Lopes et al., 2018 | Pentacam HR, OCULUS GmbH, Wetzlar, Germany software (version 1.20r118) | Random forest, SVM, neural network, regularized discriminant analysis, Naive Bayes | ||
| Silverman et al., 2014 | Artemis-1 (ArcScan, Inc., Morrison, CO) very high-frequency (VHF) digital ultrasound system; Procyon P2000 pupillometer (Haag-Streit, Bern, Switzerland). Tomography was assessed using the Orbscan II (Bausch & Lomb, Claremont, CA), and topography and simulated keratometry (K) were assessed using the Atlas corneal topography system (Carl Zeiss Meditec AG, Dublin, CA). | Neural network | ||
| Ruiz Hidalgo et al., 2016 | Pentacam HR, software version 1.20 r02 (Oculus, Wetzlar, Germany) | SVM | ||
| Souza et al., 2010 | Orbscan IITM (Bausch & Lomb) | SVM, neural network | ||
| Saika et al., 2013 | TMS-4 Advance Corneal Topographer (Tomey Corporation, Nagoya, Japan) corneal topographer (KR-9000PW, TOPCON, Tokyo, Japan) | kNN, linear discriminant, Mahalanobis distance (DIS), neural network | ||
| Lopes et al., 2015 | Pentacam HR, OCULUS GmbH, Wetzlar, Germany | SVM | ||
| Ambrósio et al., 2017 | Pentacam HR and Corvis ST (OCULUS Optikgeräte GmbH, Wetzlar, Germany) | Random forest | ||
| Yousefi et al., 2018 | SS-1000 CASIA OCT Imaging Systems (Tomey, Japan) | Density-based clustering | ||
| Lavric and Valentin, 2019 | Not mentioned in the manuscript | Neural network | ||
| Maeda et al.,1995 | Videokeratocope (TMS-1, computed anatomy, NY) | Neural network | ||
| Accardo and Pensiero, 2003 | Videokeratoscope (EyeSys) | Neural network | ||
| Feizi et al., 2016 | Tomey, EM-3000, version 4.20, Nagoya, Japan | Decision tree | ||
| Carvalho, 2005 | EyeSys System 2000 (EyeSys Vision, Houston, TX) | Neural network | ||
| Arbelaez et al., 2012 | Sirius, software version 1.2, CSO, Firenze, Italy | SVM | ||
| Ruiz Hidalgo et al., 2017 | Pentacam HR (v1.20 r53; OCULUS GmbH, Wetzlar, Germany) | SVM | ||
| Castro-Luna et al., 2000 | CSO topography system (CSO, Firenze, Italy) | Bayesian network | ||
| Kamiya et al., 2019 | CASIA SS-1000 (Tomey, Aichi, Japan) | Deep learning |
ARTIFICIAL NEURAL NETWORK
An artificial neural network (ANN) is an ML model inspired by brain architecture. The central concept behind an ANN is the neuron, which is the computation unit that acquires knowledge during the training process (using a training algorithm). The neurons are organized and composed in layers that are organized in multiple ways, resulting in several ANN architectures. Fig. 13.2 illustrates a basic example of an ANN used for KC diagnosis. The use of multiple successive layers for learning (several modern ANN architectures involve hundreds of layers) is called DL.
DECISION TREES
Decision trees are structures that divide the data based on a series of questions. Each internal node represents a question based on a feature (e.g., is the inferior-superior asymmetry lower than 1.4 diopter [D]?). Each branch of the tree is the result of the internal nodes (e.g., the answer to the previous question can be yes/no), branching out to reach the leaf nodes (which are nodes that no longer give rise to other branches). The leaf nodes are the final result of the task (e.g., in KC diagnosis, they decide if a patient eye has the condition or not) ( Fig. 13.3 ). For simplicity, imagine a tree where each branch is the product of a question until a final branch is reached that no longer divides. This last branch represents the final result that is sought (e.g., keratoconus or not). Decision trees can be created manually, but in ML, optimal decision trees are automatically generated from the data by employing algorithms. An important characteristic of decision trees is their interpretability (only for small trees), which is one of the main reasons why they are used in medical fields, especially in KC diagnosis.
SUPPORT VECTOR MACHINE
SVMs aim to determine the parent class of provided sample data by finding an optimal decision boundary (the line that separates the data into two classes). The optimal decision boundary can achieve the maximum possible separation of the data (with a process called large margin classification ), compared with other boundaries where the classes can remain close together (meaning that they would not perform as well on new instances). To find this optimal decision boundary, the data (e.g., corneal parameters) are mapped to a new high-dimensional representation. To understand this model, imagine a dividing line that allows us to separate the data into two groups. The model seeks to create this line by graphing the data in dimensions (two-dimensional [2D], three-dimensional [3D], four-dimensional [4D] … nD), which is necessary to achieve the best possible division. Fig. 13.4 shows a basic example of use of a SVM for KC diagnosis.
PARAMETERS AND DEVICES USED IN AI MODELS
The most frequently used AI devices mentioned in the literature for KC diagnosis are topographers, Scheimpflug imaging analysis, anterior segment optical coherence tomography, and devices related to corneal biomechanics (e.g., Corvis ST). Table 13.1 lists the devices used in studies of AI in KC diagnosis. The most frequently used parameters in ML models are the curvature map, corneal irregularity indices, anterior chamber parameters, and pachymetry.
SENSITIVITY AND SPECIFICITY
Three studies achieved 100% sensitivity and specificity in distinguishing normal from KC eyes: Smolek and Klyce, using an ANN approach based on topographic parameters in 33 KC eyes; Ambrósio et al., using a random forest approach in a multicenter study combining data from corneal deformation response and tomography corneal data in 204 eyes; and Castro-Luna et al., using a Naive Bayes classifier based on Placido-disk indices in 30 eyes. However, the total sample sizes (including controls and KC eyes) used in these studies were different (300, 850, and 60, respectively), as were the number of classes used, affecting their results. For example, Smolek and Klyce included 11 classes, with 66 KC eyes, whereas Ambrósio et al. included 204 eyes and four classes. Table 13.2 shows the sensitivity, specificity, accuracy, and area under the receiving operating characteristic curve (AUROC) for each study analyzing normal and KC eyes.
