Deep learning approach for analyzing regional scheimpflug corneal images to assess keratoconus progression.

PURPOSE: To use deep learning models on baseline Scheimpflug corneal images from three different corneal regions to discriminate progressive from nonprogressive keratoconus.

METHODS

A total of 562 eyes of 325 keratoconus patients were included in the study. The eyes were classified as those with progression (n = 300) and without progression (n = 262) according to the Pentacam Belin ABCD progression display (Oculus, Wetzlar, Germany). Images of the whole cornea, central 5-mm zone, and central 3-mm zone from Pentacam tomography maps (anterior and posterior axial/sagittal curvature, anterior and posterior tangential curvature, anterior and posterior elevation, corneal thickness) obtained at the patients' baseline visit were used to train and test two convolutional neural network models (YOLOv8L and YOLOv8X).

RESULTS

The central 3-mm zone of the posterior axial/sagittal curvature map (in YOLOv8L) and posterior tangential curvature map (in YOLOv8X) had the highest test accuracy in discriminating progressive from nonprogressive cases (70% [95% CI, 57-80%] for both). Sensitivity and specificity values were 74% (95% CI, 56-86%) and 68% (95% CI, 49-83%) for the posterior axial/sagittal curvature map; 92% (95% CI, 77-98%) and 63% (95% CI, 44-79%) for the posterior tangential curvature map, respectively.

CONCLUSIONS

Using the YOLOv8L and YOLOv8X deep learning models, the central 3-mm zone of the posterior axial/sagittal and posterior tangential curvature maps appeared comparatively more informative in identifying progressive keratoconus. These preliminary findings suggest that deep learning models may have the potential to distinguish keratoconus progression by analyzing the central 3-mm posterior corneal maps, although further studies incorporating clinical risk factors are needed.