Open-Source Periorbital Segmentation Dataset for Ophthalmic Applications.

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Accurate segmentation of periorbital structures is crucial for oculoplastic and craniofacial surgery.
• Existing datasets lack specificity for oculoplastic segmentation tasks.
• Development of robust deep learning models requires high-quality, specialized datasets.

Purpose of the Study:

• To create and validate a novel dataset for oculoplastic segmentation.
• To enable periorbital distance prediction using deep learning models.
• To facilitate advancements in craniofacial and oculoplastic surgical planning.

Main Methods:

• Utilized images from two open-source datasets, cropped to focus on the eye region.
• Involved segmentation of key periorbital structures (iris, sclera, lid, caruncle, brow) by five trained annotators.
• Performed rigorous inter- and intragrader reliability analyses using Dice scores.
• Trained three DeepLabV3 segmentation models to demonstrate dataset utility.

Main Results:

• Annotated a total of 2842 images with high inter- and intragrader reliability (average Dice scores of 0.82 and 0.81, respectively).
• Achieved strong performance with segmentation models, with Dice scores up to 0.90 for models trained on combined datasets.
• Demonstrated the dataset's effectiveness in training deep learning models for segmentation tasks.

Conclusions:

• Developed a unique, publicly available dataset for oculoplastic and craniofacial segmentation.
• The dataset supports the rapid development of clinically relevant segmentation networks for periorbital distance prediction.
• Provided open-source toolkits and model weights to foster community-driven innovation in the field.