Diagnosis of Subclinical Keratoconus Based on Machine Learning Techniques

Abstract

(1) Background: Keratoconus is a non-inflammatory corneal disease characterized by gradual thinning of the stroma, resulting in irreversible visual quality and quantity decline. Early detection of keratoconus and subsequent prevention of possible risks are crucial factors in its progression. Random forest is a machine learning technique for classification based on the construction of thousands of decision trees. The aim of this study was to use the random forest technique in the classification and prediction of subclinical keratoconus, considering the metrics proposed by Pentacam and Corvis. (2) Methods: The design was a retrospective cross-sectional study. A total of 81 eyes of 81 patients were enrolled: sixty-one eyes with healthy corneas and twenty patients with subclinical keratoconus (SCKC): This initial stage includes patients with the following conditions: (1) minor topographic signs of keratoconus and suspicious topographic findings (mild asymmetric bow tie, with or without deviation; (2) average K (mean corneal curvature) <46, 5 D; (3) minimum corneal thickness (ECM) > 490 μm; (4) no slit lamp found; and (5) contralateral clinical keratoconus of the eye. Pentacam topographic and Corvis biomechanical variables were collected. Decision tree and random forest were used as machine learning techniques for classifications. Random forest performed a ranking of the most critical variables in classification. (3) Results: The essential variable was SP A1 (stiffness parameter A1), followed by A2 time, posterior coma 0º, A2 velocity and peak distance. The model efficiently predicted all patients with subclinical keratoconus (Sp = 93%) and was also a good model for classifying healthy cases (Sen = 86%). The overall accuracy rate of the model was 89%. (4) Conclusions: The random forest model was a good model for classifying subclinical keratoconus. The SP A1 variable was the most critical determinant in classifying and identifying subclinical keratoconus, followed by A2 time.

Efficacy of custom-made soft keratoconus lenses on corneal aberrations and photic phenomena in patients with keratoconus: a corneal topography imaging based study

“…Although topographic systems cannot measure the wavefront directly, they can estimate corneal aberrations from corneal steepening in ectatic areas that result in a deviation of rays projected onto the retina and a deformed wavefront [26] . In eyes with KC, the increase in anterior HOA, especially in coma, is caused by the irregular steepening and protrusion of the anterior corneal surface, which can be detected with high reliability by camera-based Scheimpflug topography systems [27] . However, even with RGP lenses, residual HOAs remain on the eye, typically attributed to the posterior corneal surface [7] .…”

Automatic detection of keratoconus on Pentacam images using feature selection based on deep learning

“…In their study of 61 healthy eyes and 20 eyes with subclinical keratoconus, Castro-Luna et al classified subclinical keratoconus using the random forest model with 89% accuracy, 93% specificity, and 86% sensitivity. 15 The progression of subclinical cases to clinically observable keratoconus suggests that subclinical cases have a similar but milder corneal disorder to those with keratoconus. For this reason, it would not be wrong to say that subclinical cases in the four map have changes similar to those of keratoconus, but cannot be distinguished by the human eye.…”

“…13 In recent years, deep learning-based studies that support the clinician's decision in the detection of keratoconus disease have attracted attention. [14][15][16] Kuo et al used three different convolutional neural network (CNN) models for keratoconus screening based on corneal topographic images. Sensitivity and specificity were determined at 0.90 levels with these models.…”

“…Artificial intelligence algorithms have been used successfully in the evaluation of eye‐related conditions such as glaucoma, 11 retinal diseases, 12 and cataracts 13 . In recent years, deep learning‐based studies that support the clinician's decision in the detection of keratoconus disease have attracted attention 14–16 …”

Artificial intelligence and corneal diseases

“…Other studies have compared machine learning algorithms to detect ffKCN. Castro-Luna et al [38] found that the random forest outperformed decision tree model (89% accuracy vs. 71%, respectively) based on tomographic and biomechanical variables. Cao et al [39] also found the random forest model outperformed other machine learning algorithms using tomographic and demographic data, while Aatila et al [40] found the random forest model to have the highest accuracy compared to other machine learning models trained on anterior segment (AS)-OCT images in detecting all classes of KCN including ffKCN.…”