Fig 1.
(A) Appearance of the device. (B) Optical path diagram. (C) Principle of interference. (D) Example of interference color image. (E) Example of interference color image with tear breakup.
Fig 2.
Examples of image patches for each class of training data.
(A) Tear breakup classes. (B) Non-breakup classes.
Table 1.
Number of images for each class of training data.
Table 2.
Detailed breakdown of the six groups of training data.
Table 3.
Combination of groups for six-fold cross-validation.
Fig 3.
Method for tear breakup detection in a frame image.
The target region for detection (384 x 384 pixels) is divided into 96 x 96 pixels segments with a stride length of 48 pixels, and each segment region is classified by the trained model. When the number of segments classified as breakup is equal to or greater than a certain number, the frame image is determined to have tear breakup.
Fig 4.
(A) Confusion matrix for image classification of test data in six-fold cross validation. (B) Confusion matrix aggregated for breakup and non-breakup.
Fig 5.
Receiver operating characteristic (ROC) curve.
This curve shows that our method achieved an area under curve (AUC) of 0.898, a sensitivity of 84.3%, and a specificity of 83.3% in detecting tear breakup for a frame image.
Fig 6.
Each cell shows the ratio of the number of predicted data to the number of actual data.
Fig 7.
Examples of misclassified images.
(A) Line pattern misclassified to Spot pattern. (B) Spot pattern misclassified to Line pattern. (C) Breakup patterns misclassified to Particle: (i) Spot pattern and (ii) Line pattern.
Fig 8.
Examples of a frame image where breakup was detected correctly.
The colored regions are where tear breakup was detected: (A) Area pattern (purple), (B) Spot pattern (blue), and (C) Line pattern (green).
Fig 9.
Examples of a frame image where detection of breakup failed.
In (A) and (B), the breakup detection area was too small to determine the presence of breakup: (A) Spot pattern and (B) Line pattern. In (C), the high intensity reflections along the lower eyelid were incorrectly detected as Line pattern breakup.