Pixel-Level Segmentation of Retinal Breaks in Ultra-Widefield Fundus Images with a PraNet-Based Machine Learning...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal breaks are critical lesions that can lead to vision loss if not treated promptly.
• Accurate detection of retinal breaks in ultra-widefield fundus (UWF) images is challenging for early intervention.

Purpose of the Study:

• To develop and validate a deep learning segmentation model for localizing retinal breaks in UWF images.
• To assess the performance of a PraNet-based model for pixel-level segmentation of retinal breaks.

Main Methods:

• A deep learning segmentation model utilizing the PraNet architecture was developed.
• The model was trained and evaluated on a large dataset of 34,867 UWF images from 8083 cases.
• Performance metrics included accuracy, precision, recall, IoU, dice score, and centroid distance score.

Main Results:

• The model achieved high accuracy (0.996) and robust segmentation performance.
• Key metrics: precision 0.635, recall 0.756, IoU 0.539, dice score 0.652, centroid distance score 0.081.
• This represents the first pixel-level segmentation of retinal breaks in UWF images using deep learning.

Conclusions:

• The developed PraNet-based deep learning model shows significant potential for clinical application in detecting retinal breaks.
• Accurate and automated delineation of retinal breaks can improve early diagnosis and treatment of retinal detachment.
• The model's robust performance highlights the advancement in AI for ophthalmic image analysis.