Related Experiment Video
Updated: Mar 6, 2026
Intense Pulsed Light for the Treatment of Dry Eye Owing to Meibomian Gland Dysfunction
Published on: April 1, 2019
Machine Learning-Based Classification and Dynamic Analysis of Tear Film Lipid Layer Using Smartphone-Based
Yoshiro Okazaki1, Hiromichi Okazaki, Mamoru Iwabuchi
1Faculty of Human Sciences (Y.O., H.O., M.I.), Waseda University, Tokorozawa, Japan; and Department of Ophthalmology (N.Y.), Kyoto Prefectural University of Medicine, Kyoto, Japan.
Objectives:
To develop and validate a machine learning (ML) model for classifying tear film lipid layer (TFLL) patterns from self-acquired smartphone-based interferometer (SBI) images and evaluate its applicability to dynamic TFLL monitoring outside clinical settings.
Methods:
A healthy female participant captured TFLL videos using SBI on four days (days 0, 1, 2, and 5). RGB images were extracted and annotated into four classes: Colorful, Grayish, Transparent, and Nonregion of interest. A total of 89,033 patches were used to train ML model with Lab color and gray level co-occurrence matrix texture features. All-days fold cross-validation and Last-day fold validation schemes were used. The dynamic behavior of TFLL postblink was assessed by comparing time-series changes in TFLL area detected manually and by the ML model.
Results:
Accuracy and macro-F1 were 0.853 and 0.755 in all-days fold, and 0.829 and 0.587 in last-day fold. Machine learning-based TFLL-area estimates correlated strongly with manual measurements (r=0.969, P <0.001), capturing consistent postblink expansion.
Conclusions:
The proposed ML model can classify TFLL patterns from self-acquired SBI images with good accuracy and replicate typical TFLL spreading dynamics. These findings support its potential application to dynamic monitoring in personalized eye care and home-based dry eye management.
