Keratoconus detection of changes using deep learning of colour-coded maps.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Keratoconus is a progressive eye condition affecting corneal shape.
• Accurate detection and grading are crucial for timely management.
• Current diagnostic methods can be subjective or time-consuming.

Purpose of the Study:

• To assess the accuracy of convolutional neural networks (CNNs) in detecting keratoconus.
• To evaluate CNN performance using colour-coded corneal maps from Scheimpflug imaging.
• To determine if CNNs can grade keratoconus severity.

Main Methods:

• A multicentre retrospective study utilized corneal tomography scans from keratoconic and healthy individuals.
• Convolutional neural network (CNN) models were trained and tested using axial, anterior/posterior elevation, and pachymetry maps.
• An independent testing set from a different center was used for validation.

Main Results:

• The CNN model achieved an accuracy of 0.9785 in detecting keratoconus versus healthy eyes when using all four corneal maps.
• Individual map accuracies ranged from 0.9283 to 0.9749.
• The model accurately differentiated between keratoconus stages with accuracies between 0.8537 and 0.9032.

Conclusions:

• Convolutional neural networks demonstrate excellent performance in detecting keratoconus.
• CNNs accurately grade keratoconus severity using Scheimpflug-derived corneal maps.
• This AI approach holds potential for keratoconus screening and patient management.