Unsupervised corneal contour extraction algorithm with shared model for dynamic deformation videos: improving...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate corneal contour extraction is crucial for biomechanical evaluation and clinical diagnosis.
• Existing methods often require manual annotation, which is time-consuming and prone to error.
• Dynamic corneal videos frequently contain noise, complicating contour detection.

Purpose of the Study:

• To develop an automatic corneal contour extraction algorithm using a shared model for dynamic corneal videos.
• To improve the accuracy and efficiency of corneal biomechanical evaluation and clinical diagnoses.
• To enable unsupervised semantic segmentation of corneal regions without manual labeling.

Main Methods:

• A fully convolutional deep-learning network was employed for unsupervised semantic segmentation.
• The algorithm utilizes corneal geometry and texture information for contour extraction.
• A shared model mechanism was used to accelerate segmentation and resist noise in videos.

Main Results:

• The algorithm was validated on 1027 corneal videos acquired using an ultra-high-speed Scheimpflug camera.
• Achieved an Intersection over Union (IoU) of 95% and an average overlap error of 0.05.
• Demonstrated effective noise resistance and accurate determination of corneal regions based on shape factors.

Conclusions:

• The developed algorithm automatically extracts accurate corneal contours from noisy videos without manual annotation.
• It offers good repeatability and outperforms other algorithms in accuracy and efficiency.
• This method enhances corneal biomechanical evaluation and clinical diagnoses.