Summary and Conclusions

Chapter 8: Summary and Conclusions

• Keratoconus is a relatively common condition in our country, usually affecting young people and affecting males and females equally.
• The network for the anterior tangential curvature achieved high accuracy in diagnosing definite keratoconus. The networks for the anterior and posterior elevation maps and the sagittal curvature achieved high accuracy in excluding the diagnosis.
• The networks for the anterior and posterior refractive power and the anterior and posterior tangential curvature could exclude the presence of a normal cornea with high ability.
• No map network could confirm or exclude the diagnosis of a suspect cornea. This may be because these corneas represent a spectrum between normal and keratoconic corneas and no characteristic patterns can be extracted.
• The AI system achieved high accuracy in distinguishing the three classes (keratoconic, normal, and suspect). The most accurate networks were, in this order, the network for the anterior elevation map, the posterior elevation map, the anterior refractive power, and the equivalent refractive power.
• AI system, SIRIUS device software, and physician achieved similar results; no statistically significant difference existed between them.
• The physician with AI system support achieved the best result compared with all other models, including the physician with SIRIUS software support; the difference was statistically significant.
• Examination of the heatmaps showed that the neural networks could distinguish characteristic patterns of normal and keratoconic corneas on the various maps.
• Examination of controversial cases showed that no model and no system is free of limitations. History-taking and clinical examination of both eyes are essential for correct diagnosis, particularly in borderline cases. The importance of these systems and models lies in image pre-screening and in providing rapid and accurate support to the physician in decision-making.
• The case examination also showed that the neural networks recognised irregular patterns (aberrations) on elevation maps, tangential curvature maps, and posterior curvature maps that may be compatible with definite or suspect keratoconus.