image: Standardized annotation strategies for meibomian gland imaging include gland segmentation, overall gland mapping, ghost gland exclusion, and tortuosity assessment. These methods improve annotation consistency, reduce labeling errors, and support the development of reliable AI-assisted dry eye diagnostic systems.
Credit: Intelligent Medicine. Image source link: https://www.sciencedirect.com/science/article/pii/S2667102625001378
- Inconsistent imaging data standards have constrained AI-assisted dry eye diagnosis. This consensus establishes the first systematic classification and annotation standards across five core modalities: tear film lipid layer, tear meniscus height, tear film breakup time, corneal fluorescein staining, and meibomian gland images.
- Lipid layer grading uses a three-tier color scoring scale and a seven-level interferometric classification (0 to 160 nm), with interference subtypes mapped to specific dry eye clinical subtypes for AI-assisted subtype diagnosis.
- Tear meniscus height protocols correct a common boundary error: the lower meniscus edge is the tear-eyelid skin junction, not the Placido ring projection, improving data accuracy across OCT and ocular surface analyzer platforms.
- Corneal fluorescein staining is encoded through two systems: a four-grade severity classification and a nine-point three-region scoring system, covering the full spectrum of ocular surface damage as AI-trainable labels.
- For meibomian gland imaging, ghost glands (non-secreting atrophied remnants) are excluded from positive gland annotations, correcting a systematic labeling error. In vivo confocal microscopy standards cover acinar density, diameter, inflammatory cell count, and glandular dropout.
- A quality control framework covers annotator qualifications, image preprocessing, inter-annotator consistency using the kappa coefficient, multi-round review, and data cleaning.
- Five structural challenges are identified and addressed, including uneven data quality, inconsistent annotation standards, and limited data sharing. Federated learning and privacy-preserving de-identification are recommended to enable large-scale multicenter collaboration.
Reference
Title of original paper: Expert consensus on classification and annotation methods, processes, and quality control for dry eye imaging in artificial intelligence applications (2025)
Journal: Intelligent Medicine
DOI: 10.1016/j.imed.2025.05.012
Journal
Intelligent Medicine
Article Title
Expert consensus on classification and annotation methods, processes, and quality control for dry eye imaging in artificial intelligence applications (2025)