In this study, we aimed to establish and evaluate an index to define keratoconus progression based on map images obtained from anterior segment optical coherence tomography (AS-OCT) and age through deep learning. AS-OCT images and patient data were retrospectively retrieved from Hospital records. Classifiers for each of the six types of images were created using machine learning, and an index was established for judging keratoconus progression from the output adjusted or unadjusted for age. A total dataset of 2006 paired AS-OCT images from longitudinal examinations (positive: 883, negative: 1123), comprising a learning dataset of 1678 pairs (positive: 741, negative: 937) and a validation dataset of 328 pairs (positive: 142, negative: 186), was used. The final numerical value (index) was obtained from the output results and age. The index was subsequently evaluated using the validation dataset to determine its accuracy, sensitivity, specificity, goodness of fit, and area under the receiver operating characteristic curve (AUC). The average value of the keratoconus progression index significantly differed between the age-adjusted and unadjusted algorithms (p<0.001). Adjusting for age improved the keratoconus progression index (accuracy, 0.655 vs. 0.909; sensitivity, 0.915 vs. 0.937; specificity, 0.457 vs. 0.887; precision, 0.563 vs. 0.864; F-value, 0.697 vs. 0.899; and AUC 0.727 vs. 0.935). Using machine learning on six types of AS-OCT images adjusted for age, we developed an index with a better predictive value of keratoconus progression than previously reported. This index accurately predicts keratoconus progression, which is useful for determining indications for corneal cross-linking using initial AS-OCT imaging.
Miyai et al. (Wed,) studied this question.