Development of the AI Pipeline for Corneal Opacity Detection

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Global ophthalmological services are inadequate, particularly in low- and middle-income countries, due to shortages in practitioners and equipment.
• Corneal opacity is a significant global cause of blindness, necessitating improved diagnostic methods.

Purpose of the Study:

• To detect corneal opacity from videos of the anterior eye segment.
• To develop an artificial intelligence (AI) pipeline for accurate corneal opacity detection.

Main Methods:

• Extracted video frames and processed them using a convolutional neural network (CNN).
• Manually annotated images, extracted corneal margins, and applied contrast adjustment (CLAHE).
• Performed semantic segmentation of the cornea using annotated data.

Main Results:

• Achieved an accuracy of 0.8 for raw image frames and 0.96 for processed corneal margins.
• Obtained Dice and IoU scores of 0.94 for semantic segmentation of corneal margins.
• Demonstrated significant accuracy improvement through manual annotation, corneal extraction, and CLAHE.

Conclusions:

• AI pipeline with manual annotation and semantic segmentation achieves high accuracy in detecting corneal opacity.
• Portable slit lamp microscopy and AI offer potential for equitable global diagnostic resource distribution.
• Enhanced image processing techniques are crucial for improving AI performance in ophthalmic diagnostics.