Toward Keratoconus Diagnosis: Dataset Creation and AI Network Examination

Artificial intelligence (AI) has emerged as a prominent technology across diverse domains, including ophthalmology, where the disease-grade diagnosis is crucial. Diagnosing keratoconus (KC) is essential due to its effect on refractive surgeries like LASIK and Femto. Before performing these procedures, surgeons must thoroughly evaluate the patient’s eye health. AI can significantly assist in the automatic and accurate diagnosis of keratoconus. For this purpose, the development of a comprehensive dataset of medical images and disease severity information is essential. The present research created a dataset of 6,000 Four Maps Refractive images labeled into six classes. In addition, a two-class labeling was given for the dataset, determining whether there is a barrier to the patient’s eye surgery. Our approach contains images of each patient’s left and right eyes, emphasizing the need to scan both eyes simultaneously. This method facilitates an accurate diagnosis of keratoconus and provides an accurate representation of its temporal pattern. The previous datasets did not examine the consequences of concurrently assessing both eyes and used a smaller number of classes to categorize the data. Additionally, this study has implemented the transfer learning approach with convolutional neural networks to diagnose two classes: ”Non-suitable for refractive surgery” and ”Suitable for refractive surgery”. Our results demonstrate that the ResNet18 model, when used on the Four-Maps dataset, has an accuracy rate of 94%. It has also been demonstrated that when only one of the color maps (Sagittal, Thickness, Elevation Front, and Elevation Back) is available, the EfficientNetB0 model performs better. It achieved 94%, 93%, 94%, and 92% accuracy for each map, respectively.