Open-Source Periorbital Segmentation Dataset for Ophthalmic Applications

TL;DR

This work addresses the lack of open-source, clinically detailed periorbital segmentation data capable of supporting sub-millimeter distance prediction. The authors assemble two open-source datasets (CFD and CelebAMask-HQ) with comprehensive annotations for iris, sclera, lids, caruncle, and brow, and release a distance-prediction toolkit and pretrained weights. Segmentation experiments show strong performance for iris, sclera, and brow (Dice 0.78–0.96) with more limited performance for caruncle and lids, reflecting annotation challenges and dataset differences; inter- and intra-grader reliability confirms annotation quality while highlighting inherent subjective variability. By providing open data, tools, and benchmarks, the work facilitates rapid development of clinically useful periorbital segmentation networks and distance-prediction methods that can support disease monitoring and remote ophthalmic care.

Abstract

Periorbital segmentation and distance prediction using deep learning allows for the objective quantification of disease state, treatment monitoring, and remote medicine. However, there are currently no reports of segmentation datasets for the purposes of training deep learning models with sub mm accuracy on the regions around the eyes. All images (n=2842) had the iris, sclera, lid, caruncle, and brow segmented by five trained annotators. Here, we validate this dataset through intra and intergrader reliability tests and show the utility of the data in training periorbital segmentation networks. All the annotations are publicly available for free download. Having access to segmentation datasets designed specifically for oculoplastic surgery will permit more rapid development of clinically useful segmentation networks which can be leveraged for periorbital distance prediction and disease classification. In addition to the annotations, we also provide an open-source toolkit for periorbital distance prediction from segmentation masks. The weights of all models have also been open-sourced and are publicly available for use by the community.

Figures (7)

• Figure 1: Representative images and annotations from the CFD and Celeb dataset used to construct the dataset described here.
• Figure 2: Schematic of preprocessing and training pipeline. Full details are described in the methods, but briefly, the dataset was split using an 80/20 train test split. The input image was split at the midline, and both halves of the image (and label) were resized to 256x256. A DeepLabV3 model with a ResNet101 backbone pretrained on ImageNet1K was trained for 500 steps. The same preprocessing procedure was used at test time. Following segmentation, the left and right halves of the image were resized and stitched back together such that the full segmentation mask was the same size as the input.
• Figure 3: Pairwise matrices representing intergrader agreement as the average Dice score between graders or DeepLabV3 over 100 randomly sampled images. A) The average pairwise Dice score between all graders, and B-F) represent the Dice score on the iris, sclera, brow, lid, and caruncle classes respectively.
• Figure 4: Periorbital distances on two images from the CFD dataset. These distances can be calculated using the toolkit, which we have made available via API, and the periorbital distances from the CFD dataset have been released as a benchmark dataset.
• Figure 5: Examples of how different types of lids were annotated using the lid crease as a guide. A) No visible lid crease, so no annotation was created. B) When the lid crease was present, the lid annotation included any visible lashes so the boundaries of the masks align with the sclera mask C) In the event of a partially visible lid crease, only regions of the lid inferior to the lid crease were annotated.