Efficacy of a Deep Learning Convolutional Neural Network (CNN) in the Early Detection and Automated Grading of Subclinical Keratoconus Using Corneal Topography
- May 18
- 2 min read
DOI: https://doi.org/10.66715/jsccr/2026v3.i4.17 | Original Research | 2026 | Volume 3 | Issue 4 | Page 1-7
Ankit Sharma, M. Optom, PhD (Pursuing), Assistant Professor, Om Sterling Global University, Hisar, Email- aksharma90534@gmail.com
Abstract
Background and Objective:
Keratoconus (KC) is a progressive corneal thinning disorder that leads to irregular astigmatism and visual impairment. Detecting the disease at its subclinical stage remains a significant clinical challenge, as these patients often present with normal visual acuity and no distinct signs on slit-lamp examination, posing a severe risk for post-refractive surgery ectasia. This study aims to develop and evaluate the efficacy of a Deep Learning Convolutional Neural Network (CNN) for the early detection and automated grading of subclinical keratoconus using corneal topography maps.
Methods:
In this retrospective study, a dataset of 1,250 corneal topography images—including Axial Curvature, Anterior/Posterior Elevation, and Pachymetry maps—was utilized. The dataset comprised normal eyes, subclinical keratoconus, and advanced keratoconus cases. A customized CNN architecture (based on a ResNet-50 backbone) was trained and evaluated using an 80:10:10 split for training, validation, and testing, respectively. Model performance was assessed using Sensitivity, Specificity, and the Area Under the Receiver Operating Characteristic curve (AUROC).
Results:
The proposed CNN model achieved an overall accuracy of 96.4% in differentiating subclinical keratoconus from normal eyes, demonstrating a Sensitivity of 95.2%, a Specificity of 97.6%, and an AUROC of 0.988. For automated severity grading (Stages 1 to 4), the model achieved a classification accuracy of 91.8%. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed that the network accurately localized features corresponding to localized corneal thinning and steepest curvature.
Conclusion:The deep learning CNN model demonstrates high efficacy in the early detection and precise automated grading of subclinical keratoconus from corneal topography. This automated system holds strong potential as a clinical decision-support tool to streamline screening processes and enhance patient safety in refractive surgery clinics.
Keywords: Keratoconus, Subclinical Keratoconus, Corneal Topography, Deep Learning, Convolutional Neural Network (CNN), Automated Grading.