Prediction of ectatic corneal diseases from raw optical coherence tomography data by using convolutional neural networks

Abstract

Purpose The purpose of this study is to use three-dimensional volumetric raw optical coherence tomography data to predict the ectasia screening index, a metric derived from the Casia2 anterior-segment optical coherence tomographer (Tomey, Nagoya, Japan) for detecting ectatic corneal diseases, and to classify the corneal condition through the application of convolutional neural networks. Methods The CNNs DenseNet121, EfficientNet-B0, and ResNet18 were modified for ESI prediction. Additionally, DenseNet121, EfficientNet-B0, MobileNetV3-Large, and ResNet18 were modified to classify the corneal condition into three categories: ‘normal’, indicating a healthy eye; ‘ectasia’, indicating corneal ectasia; and ‘other disease’, which comprised eyes with corneal dystrophies, penetrating keratoplasty, Salzmann's nodules, pterygium, subepithelial or stromal scarring. Moreover, EfficientNet-B0 was modified to mimic the ESI using a dataset of 4,898 training files, 620 validation files, and 623 test files, each containing 16 equiangular meridional images. Two-class and three-class classification models were evaluated: the two-class model distinguished no detectable ectasia from suspected ectasia or clinical ectasia, while the three-class model separated no detectable ectasia, suspected ectasia, and clinical ectasia. The performance of the CNN architectures was assessed using measures including accuracy, F1 score, positive predictive value, sensitivity, and specificity based on data from patients who were examined at the Department of Ophthalmology, Saarland University Medical Centre, Homburg, Germany. Results The adapted ResNet18, DenseNet121, and EfficientNet-B0 achieved accuracies of 94.80%, 95.27%, and 95.83%, respectively; F1 scores of 93.38%, 94.09%, and 94.72%, respectively; positive predictive values of 94.72%, 93.55%, and 95.28%, respectively; sensitivities of 92.07%, 94.64%, and 94.17%, respectively; and specificities of 96.61%, 95.69%, and 96.92%, respectively. For the classification of corneal condition, the modified DenseNet121, modified EfficientNet-B0, modified MobileNetV3-Large, and modified ResNet18 attained the overall accuracy of 91.27%, 91.27%, 92.86%, and 89.68%, respectively. The modified DenseNet121, modified EfficientNet-B0, modified MobileNetV3-Large, and modified ResNet18 achieved macro-averaged sensitivities of 91.27%, 91.27%, 92.86%, and 89.68%; macro-averaged specificities of 95.63%, 95.63%, 96.43%, and 94.84%; macro-averaged positive predictive values of 91.58%, 91.65%, 92.91%, and 90.24%; and macro-averaged F1 scores of 91.35%, 91.29%, 92.85%, and 89.81%, respectively. When evaluated on the 623 test files, each containing 16 equiangular meridional images, the modified EfficientNet-B0 recorded a mean absolute error of 6.65. In the two-class classification task, the architecture attained an accuracy of 87.96%, an F1 score of 89.33%, a positive predictive value of 97.52%, a sensitivity of 82.41%, and a specificity of 96.69%. In the three-class classification task, it achieved an overall accuracy of 84.75%, a weighted-average F1 score of 84.95%, a weighted-average sensitivity of 0.8475, a weighted-average specificity of 0.9333, and a weighted-average positive predictive value of 0.8525. Conclusion The effective deployment of convolutional neural network architectures using raw optical coherence tomography data demonstrates the viability of raw optical coherence tomography data for diagnosing ectatic corneal diseases.