Neural network-based system for early keratoconus detection from corneal topography

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Keratoconus identification, especially in early stages, poses diagnostic challenges.
• Current neural network approaches for keratoconus screening typically analyze each eye independently.
• The optimal strategy for keratoconus screening—bilateral versus unilateral analysis—remains unclear.

Purpose of the Study:

• To compare the diagnostic capability of neural networks for keratoconus screening using bilateral versus unilateral eye data.
• To optimize neural network parameters for improved keratoconus detection and classification.

Main Methods:

• Neural networks were trained and tested using various configurations of input/hidden/output nodes and learning rates.
• Input parameters included data from both eyes of a subject, and output represented three clinical categories (normal, keratoconus, other).
• The study explored different combinations of network architecture and learning rates to find optimal settings.

Main Results:

• The best performing model utilized parameters from both eyes as input and classified subjects into three categories.
• Optimal parameters included a low number of hidden layer neurons (less than 10) and a learning rate of 0.1.
• This approach achieved a global sensitivity of 94.1% (100% for keratoconus) and a global specificity of 97.6% (98.6% for keratoconus).

Conclusions:

• Screening for keratoconus using neural networks is more effective when considering data from both eyes simultaneously.
• The developed method demonstrates high accuracy in identifying keratoconus and differentiating it from other corneal conditions.
• This bilateral approach offers a promising advancement for early diagnosis and effective screening of keratoconus.