Detection Of Keratoconus Region Of Interest In Topographic Maps Using Graphs

Early diagnosis and monitoring of keratoconus (KC) are essential. Current methods often lack sufficient sensitivity and specificity to detect subtle corneal changes. This study introduces a methodology using graph-based algorithms on corneal topographical maps to identify and localize the cone area more effectively. Specifically, it aims to evaluate and compare this graph-based approach against conventional topographic metrics in patients with similar anterior curvature values.

Hospital de la Princesa, Madrid, Spain

A total of 211 corneal topographies (114 KC and 97 Non-KC) were analyzed using an MS-39 (CSO) tomographer. The KC group included mild cases according to the ABCD classification (stages 0–I–II), whereas the Non-KC group exhibited high physiological astigmatism but no signs of keratoconus.

The conventional metrics included the thinnest corneal thickness and anterior and posterior curvature.

Posterior surface elevation was used to calculate divergence values for identifying locally significant points in the cone zone. These points served as nodes in a graph.

Cross-validation was utilised to compute sensitivity, specificity, and accuracy, while the McNemar test was employed to assess the statistical significance of differences between the models.

The graph-based method achieved a mean sensitivity of 0.88 ± 0.09, a specificity of 0.91 ± 0.04, and an accuracy of 0.89 ± 0.06, outperforming standard topographic metrics (sensitivity: 0.85 ± 0.09; specificity: 0.77 ± 0.15; accuracy: 0.80 ± 0.06). The McNemar test confirmed statistically significant differences in overall classification (test statistic = 13.0, p = 0.04) and specificity (test statistic = 5.0, p = 0.04) when comparing the graph-based approach (M1) to the conventional metrics model (M2).

Graph-based construction provides greater accuracy and specificity than conventional topographic metrics, demonstrating robustness against artifacts outside the area of maximal curvature. Graph-based methods offer an enhanced representation of corneal structure by capturing spatiotemporal information and relationships between selected points. This enhanced representation has potential applications in both diagnosis and disease monitoring. The McNemar test further validates the statistically significant superiority of this technique. Limitations include the lack of subclinical or highly advanced cases, a need for broader population heterogeneity, and validation in different centers and devices.