A deep learning approach for keratoconus detection using spatio-temporal features from corneal imaging

Keratoconus is a progressive corneal disease that requires early and accurate detection to prevent severe visual impairment. This study presents a deep learning-based classification model for distinguishing between healthy and keratoconic eyes using dynamic corneal imaging data from the CORVIS system. A hybrid CNN-RNN architecture was developed, combining a fine-tuned InceptionV3 network for spatial feature extraction with a recurrent LSTM module to capture temporal patterns across image sequences. To ensure robust evaluation, a 10-fold stratified cross-validation strategy was employed, with data splits performed at the patient level to avoid data leakage. The model achieved an average accuracy, precision, recall, and F1-score of approximately 0.90 across folds, demonstrating strong generalization performance. Boxplot visualizations of metric distributions further confirmed model stability and revealed minimal performance variance. Class-wise analysis showed high effectiveness in detecting both healthy and keratoconic cases, although slightly greater variability was observed in the classification of healthy eyes. These results indicate that the proposed method is a promising tool for keratoconus screening and may complement existing diagnostic workflows. Further validation on external datasets is recommended prior to clinical deployment.