TearNET: Validation of a convolutional neural network for grading of tear ferning patterns using deep learning

Cont Lens Anterior Eye. 2026 Apr 11;49(3):102654. doi: 10.1016/j.clae.2026.102654. Online ahead of print.

ABSTRACT

PURPOSE: Tear-ferning patterns are microscopic crystallization formed when the solvent in the tears evaporates, and the solutes congregate to form a fern-like crystal structure. The ferning patterns are affected by the biomolecular properties of the tear and can be used to screen for dry eye disease (DED). However, manual grading of tear-ferning patterns based on Rolando's grading system is subjective and inconsistent. This study aims to validate TearNET - a convolutional neural network-based deep learning algorithm, for the automated grading of tear-ferning patterns.

METHODS: Tear samples from 80 healthy participants (160 eyes) were collected, and the ferning patterns were imaged under a microscope. Two examiners independently graded the samples, confirming grading reliability. The age and gender of participants were recorded, along with temperature and humidity in the lab during sample collection. The TearNET model was trained on these samples using a 70% training and 30% testing data split. The model's sensitivity, specificity, recall, and F-scores were estimated for performance evaluation.

RESULTS: Tear ferning patterns varied significantly with age (p < 0.001) and gender (p < 0.001). Variation in room temperature and humidity showed no significant variation (p > 0.05) on the tear ferning pattern. The TearNET model achieved 81% accuracy in classifying patterns, with notable performance for Types 3 and 4 grades.

CONCLUSIONS: Overall, the TearNET model demonstrates promising performance in classifying tear ferning patterns according to Rolando's grade, particularly excelling in differentiating individual grade patterns. The model demonstrated effective training convergence, validating its potential for future clinical applications in dry eye disease screening.

PMID:41967396 | DOI:10.1016/j.clae.2026.102654