Keratoconus diagnosis based on dynamic corneal imaging using 3D convolutional neural networks

Abstract

Keratoconus is a progressive disease that requires precise and rapid diagnosis, as well as the initiation of treatment, to prevent serious and permanent visual impairment. This article presents a comparison of 3D convolutional neural network models for the diagnosis of keratoconus based on dynamic corneal imaging results obtained with the CORVIS device. The article describes the data preprocessing and compares models of varying complexity in terms of accuracy, inference time, number of parameters, and GPU memory usage. To ensure adequate generalization capability during algorithm training, 5-fold stratified cross-validation was used, and the average metrics from all splits were compared. The best model achieved an average keratoconus detection accuracy exceeding 88%, confirming that deep neural networks can be a promising tool to support physicians in diagnosing corneal diseases such as keratoconus. Future work includes plans to gather a larger patient database and apply more advanced preprocessing methods for the video data.