AI-Driven Keratoconus Detection: Integrating Medical Insights for Enhanced Diagnosis

The proper diagnosis of keratoconus as a main ocular disorder is imperative for reducing the risk of vision blurring and potential blindness. Additionally, keratoconus (KC) significantly increases the risk of refractive eye surgeries. By using Artificial Intelligence (AI) and machine vision for early and automatic diagnosis, the risks and costs in medical systems could be significantly reduced. This study introduces an image dataset acquired by the Pentacam device, which is commonly used in keratoconus diagnosis. These images include four eye topography maps and have been labeled by three experts in two categories Suitable for refractive surgery and Non-suitable for refractive surgery. Distinguishing between these two categories in certain images poses a significant challenge. In the diagnosis of keratoconus, ophthalmologists perform a comprehensive evaluation of corneal topography maps and assign different levels of significance to each map. This research paper presents a novel algorithm that utilizes AI to imitate this medical insight. For this purpose, a regression network has been integrated into the classification algorithm to obtain the importance degree for each map. The degree of importance, which was created for each of the four maps, quantifies the level of attention given to each map in order to facilitate ideal classification. The suggested algorithm in this article utilizes a range of backbones as well as transfer learning techniques. The most favorable outcome was observed while utilizing the VGG backbone in the algorithm, yielding an F1-score of 95%, which is very promising.