Computer aided diagnosis for suspect keratoconus detection.

Purpose: To develop a stable and low-cost computer aided diagnosis (CAD) system for early keratoconus detection for clinical use.

Methods: The CAD combines a custom-made mathematical model, a feedforward neural network (FFN) and a Grossberg-Runge Kutta architecture to detect clinical and suspect keratoconus. It was applied to retrospective data of 851 subjects for whom corneal elevation and thickness data was available. These data were divided into four groups: a control group (312 eyes) with bilateral normal tomography, keratoconus suspect (77 eyes) with a clinically diagnosed keratoconus in one eye and a normal fellow eye, mild keratoconus (220 eyes), and moderate keratoconus (229 eyes). The proposed framework is validated using 10-cross-validation, holdout validation and ROC curves.

Results: The CAD detects suspect keratoconus with an accuracy of 96.56% (sensitivity 97.78%, specificity 95.56%) versus an accuracy of 89.00% (sensitivity 83.00%, specificity 95.00%) for Belin/Ambrosio Deviation (BADD), and an accuracy of 79.00% (sensitivity 58.00%, specificity 99.70%) for Topographical Keratoconus Classification (TKC). For the detection of mild to moderate keratoconus CAD shows nearly similar accuracies as previously described methods, with an average accuracy of 99.50% for CAD, versus 99.46% for BADD and 96.50% for TKC. The proposed algorithm also provides a 70% reduction in computation time, while increasing stability and convergence with respect to traditional machine learning techniques.

Conclusion: The proposed algorithm is highly accurate and provides a stable screening platform to assist ophthalmologists with the early detection of keratoconus. This framework could potentially be set up for any Scheimpflug tomography system.