Evaluation of machine learning classifiers in keratoconus detection from orbscan II examinations

Abstract

PURPOSE:To evaluate the performance of support vector machine, multi‐layer perceptron and radial basis function neural network as auxiliary tools to identify keratoconus from Orbscan II maps.METHODS:A total of 318 maps were selected and classified into four categories: normal (n = 172), astigmatism (n = 89), keratoconus (n = 46) and photorefractive keratectomy (n = 11). For each map, 11 attributes were obtained or calculated from data provided by the Orbscan II. Ten‐fold cross‐validation was used to train and test the classifiers. Besides accuracy, sensitivity and specificity, receiver operating characteristic (ROC) curves for each classifier were generated, and the areas under the curves were calculated.RESULTS:The three selected classifiers provided a good performance, and there were no differences between their performances. The area under the ROC curve of the support vector machine, multi‐layer perceptron and radial basis function neural network were significantly larger than those for all individual Orbscan II attributes evaluated (p<0.05).CONCLUSION:Overall, the results suggest that using a support vector machine, multi‐layer perceptron classifiers and radial basis function neural network, these classifiers, trained on Orbscan II data, could represent useful techniques for keratoconus detection.

Current state and future prospects of artificial intelligence in ophthalmology: a review

“…Topography and tomography provide a wealth of complex data to the ophthalmologist for each cornea. 25 Importantly, the AUROC of the MLCs were significantly larger than those obtained when evaluating each attribute individually for detecting corneal diseases. 24 A decision is often down to each individual ophthalmologist's subjective interpretation of patterns or empiric cutoff values, which vary from machine to machine.…”

“…22,23 However, identifying sub-clinical corneal ectasia remains immensely challenging. 25 Souza et al used Orbscan IIz data and tested multiple different forms of machine learning classifiers (MLC) (support vector machine, multiple layer perceptron classifiers and radial basis function neural network); all were proficient in detecting the aforementioned corneal abnormalities, with no significant difference found between their performances (Area Under the Curve of the Receiver Operating Characteristic [AUROC]: 0.98-0.99; sensitivity 0.98-1.00; specificity: 0.98-1.00). Despite this, it is very difficult for an ophthalmologist to differentiate between normal and sub-clinical keratoconus in most of the parameters analysed.…”

“…In earlier studies, results from the Orbscan IIz were considered to be optimal to data from alternative devices as posterior topography data in addition to the standard anterior information were provided. 25 Souza et al used Orbscan IIz data and tested multiple different forms of machine learning classifiers (MLC) (support vector machine, multiple layer perceptron classifiers and radial basis function neural network); all were proficient in detecting the aforementioned corneal abnormalities, with no significant difference found between their performances (Area Under the Curve of the Receiver Operating Characteristic [AUROC]: 0.98-0.99; sensitivity 0.98-1.00; specificity: 0.98-1.00). 25 Importantly, the AUROC of the MLCs were significantly larger than those obtained when evaluating each attribute individually for detecting corneal diseases.…”

“…25 Souza et al used Orbscan IIz data and tested multiple different forms of machine learning classifiers (MLC) (support vector machine, multiple layer perceptron classifiers and radial basis function neural network); all were proficient in detecting the aforementioned corneal abnormalities, with no significant difference found between their performances (Area Under the Curve of the Receiver Operating Characteristic [AUROC]: 0.98-0.99; sensitivity 0.98-1.00; specificity: 0.98-1.00). 25 Importantly, the AUROC of the MLCs were significantly larger than those obtained when evaluating each attribute individually for detecting corneal diseases. Smadja et al 23 and Hidalgo et al 24 both found their different MLCs to be highly sensitive and specific in their ability to distinguishing keratoconus from 23,24 Ultimately, these MLCs are similar in differentiating capabilities to the specifically designed non-AI indexes such as the KISA% index, 26 the Klyce/Maeda Keratoconus Index 27 and the Cone Location and Magnitude Index.…”

Evaluation of the anterior chamber angle in keratoconus and normal subjects

“…To be included, a KC subject must manifest one or more of the following clinical signs: posterior stress lines (Vogt striae), Fleischer ring, external sign (Munson sign) together with a topography positive for KC (central corneal power superior to 48.7D, and an inferior superior asymmetry above 1.9 [17][18][19]). Exclusion criteria included: any previous ocular surgery, the use of any systemic or ocular medications and any chronic disorder that can affect the eye, currently being pregnant or a nursing mother, and participating in an ophthalmologic drug or device research study within 30 days prior to entering the present study.…”