Corneal biomechanics and deep learning: A new approach to keratoconus detection

A hybrid deep learning model combining convolutional and recurrent neural network architectures applied to dynamic corneal imaging sequences achieved approximately 90% accuracy in distinguishing healthy eyes from eyes with keratoconus, according to a study published in Scientific Reports.¹ The model, developed by Syta and colleagues at institutions in Lublin, Poland, uses video sequences from the Corvis ST biomechanical imaging device and was evaluated using 10-fold stratified cross-validation on a single-center data set of 104 eyes.

Background and rationale

Traditional keratoconus diagnostic methods based on corneal topography and tomography have limited effectiveness in detecting early-stage disease.¹ The Corvis ST (Corneal Visualization Scheimpflug Technology; Oculus Optikgerate GmbH, Wetzlar, Germany) records dynamic corneal deformation induced by an air pulse using a high-speed Scheimpflug camera capturing 4,300 high-resolution images per second, generating a video sequence that captures successive phases of corneal deformation from initial state through maximum concavity and recovery.¹ The authors proposed that analyzing both the spatial and temporal features of these sequences using deep learning could improve keratoconus classification compared with approaches based on static imaging or numerical biomechanical indices alone.

Study design and data set

The data set comprised 104 Corvis ST video sequences from patients seen at the Department of Ophthalmology, Medical University of Lublin, between March and August 2024: 57 eyes with confirmed keratoconus and 47 healthy control eyes. Each video contained 139 frames at 200x576 pixels. Healthy eyes were required to have normal corneal tomography and topography results (K max < 47 D), normal elevation maps, central corneal thickness greater than 480 μm, no corneal scarring, and no family history of keratoconus. Eyes with keratoconus were required to show abnormal topography (K max > 47 D), abnormal elevation maps, and central or inferior corneal protrusion. Exclusion criteria included other corneal ectasias, corneal endothelial diseases, prior ocular surgery, and eye infections.

Preprocessing included removal of patient identifiers, device watermarks, and text overlays using a Canny edge detection and image inpainting pipeline. Canny edge detection was then applied to each video frame to extract corneal contours as diagnostically relevant features. Frames were resized to 200x200 pixels and normalized. The feature-extraction model was trained on a 55-frame segment corresponding to the main corneal deformation phase, while the recurrent component was trained on the complete 139-frame sequence. A 10-fold stratified cross-validation strategy with splits performed at the patient level was used to prevent data leakage.

Model architecture

The proposed architecture is a hybrid convolutional neural network–recurrent neural network (CNN-RNN) model. Spatial features were extracted using a fine-tuned InceptionV3 network pre-trained on ImageNet, with the last 20 layers additionally fine-tuned on each training fold to adapt the extractor to Corvis ST imaging. Temporal modeling was performed using a single long short-term memory (LSTM) layer with 16 units, followed by a dropout layer, a densely connected layer, and a sigmoid classification output. The architecture was selected over 3D CNN and Transformer-based alternatives as a compromise between representational capacity and model complexity, particularly given the limited data set size. No data augmentation or class-balancing techniques were applied.

Results

The full preprocessing pipeline, including edge detection, substantially improved performance over the model without edge detection: accuracy increased from 0.77 ± 0.19 to 0.91 ± 0.11, precision from 0.75 ± 0.20 to 0.91 ± 0.12, recall from 0.73 ± 0.25 to 0.88 ± 0.19, F1-score from 0.73 ± 0.23 to 0.88 ± 0.14, and AUC from 0.83 ± 0.17 to 0.95 ± 0.07.¹

The proposed CNN-RNN model outperformed both comparison architectures. The 3D CNN model achieved accuracy of 0.84 ± 0.09 and AUC of 0.85 ± 0.13, while the Transformer-based ViViT model achieved accuracy of 0.75 ± 0.09 and AUC of 0.74 ± 0.11, with substantially higher variability across folds. Statistical comparison using the Wilcoxon signed-rank test showed that the CNN-RNN model achieved significantly higher AUC than both the 3D CNN (P = .027) and the Transformer model (P = .004), and significantly higher accuracy (P = .016) and F1-score (P = .012) than the Transformer model.

Class-wise analysis showed that both the healthy and keratoconus classes achieved median F1-scores of approximately 0.90 to 0.92. However, the healthy class showed greater variability, including a notable outlier near 0.50 in one fold, suggesting that the model may be less stable when classifying healthy eyes in some cases, potentially due to class imbalance or overlapping feature distributions.

Limitations

The authors identified 4 key limitations. First, the data set was small (104 eyes) and single-center, which may not reflect the full spectrum of clinical variability across keratoconus stages, comorbidities, and imaging conditions. Second, greater variability in classifying healthy eyes raises the potential for false-positive diagnoses in screening settings. Third, the model has not been validated on independent external data sets or across multiple institutions. Fourth, as with most deep learning models, the current approach has limited clinical interpretability, though the preprocessing strategy constrains input to clinically relevant deformation patterns. The authors note that the pipeline was designed specifically for Corvis ST recordings, which limits the immediate availability of compatible external data sets for validation.

The authors identify external multi-center validation, expansion to include subclinical keratoconus cases, incorporation of multimodal data (topography and tomography in addition to biomechanics), and implementation of explainable AI methods such as Grad-CAM as priorities for future development. Real-time integration with diagnostic devices is also described as a long-term goal.

Reference

1. Syta A, Chorągiewicz T, Gęca J, et al. A deep learning approach for keratoconus detection using spatio-temporal features from corneal imaging. Sci Rep (2026). https://doi.org/10.1038/ s41598-026-56383-y