A texture-aware U-Net for identifying incomplete blinking from eye videography.

Area of Science:

• Ophthalmology and Computer Vision
• Medical Imaging Analysis

Background:

• Accurate identification of incomplete blinking from eye videography is crucial for diagnosing eye disorders such as dry eye.
• Current methods may lack the precision needed for subtle texture analysis in eye imaging.

Purpose of the Study:

• To develop a novel texture-aware neural network (TAU-Net) for precise palpebral fissure segmentation in eye videography.
• To enhance the early detection of incomplete blinking and related eye conditions.

Main Methods:

• Developed TAU-Net, a U-Net based convolutional neural network incorporating specialized convolutional blocks.
• These blocks utilize element-wise subtraction to emphasize subtle textures crucial for palpebral fissure segmentation.
• Trained and evaluated the network on 1396 frame images from eye videography.

Main Results:

• TAU-Net achieved a high average Dice index of 0.9587 for palpebral fissure segmentation.
• The network demonstrated a low Hausdorff distance (HD) of 4.9462 pixels.
• TAU-Net outperformed standard U-Net and its variants in segmentation accuracy.

Conclusions:

• The developed TAU-Net shows significant promise for accurate palpebral fissure segmentation in eye videography.
• This technique offers a robust approach for identifying incomplete blinking, aiding in the early diagnosis of eye diseases.
• The texture-aware approach enhances the reliability of automated eye-tracking analysis.