Non-invasive and objective tear film breakup detection on interference color images using convolutional neural networks.

PURPOSE: Dry eye disease affects hundreds of millions of people worldwide and is one of the most common causes for visits to eye care practitioners. The fluorescein tear breakup time test is currently widely used to diagnose dry eye disease, but it is an invasive and subjective method, thus resulting in variability in diagnostic results. This study aimed to develop an objective method to detect tear breakup using the convolutional neural networks on the tear film images taken by the non-invasive device KOWA DR-1α. METHODS: The image classification models for detecting characteristics of tear film images were constructed using transfer learning of the pre-trained ResNet50 model. The models were trained using a total of 9,089 image patches extracted from video data of 350 eyes of 178 subjects taken by the KOWA DR-1α. The trained models were evaluated based on the classification results for each class and overall accuracy of the test data in the six-fold cross validation. The performance of the tear breakup detection method using the models was evaluated by calculating the area under curve (AUC) of receiver operating characteristic, sensitivity, and specificity using the detection results of 13,471 frame images with breakup presence/absence labels. RESULTS: The performance of the trained models was 92.3%, 83.4%, and 95.2% for accuracy, sensitivity, and specificity, respectively in classifying the test data into the tear breakup or non-breakup group. Our method using the trained models achieved an AUC of 0.898, a sensitivity of 84.3%, and a specificity of 83.3% in detecting tear breakup for a frame image. CONCLUSIONS: We were able to develop a method to detect tear breakup on images taken by the KOWA DR-1α. This method could be applied to the clinical use of non-invasive and objective tear breakup time test.

Non-Invasive Tear Break-Up Detection with the Kowa DR-1α and Its Relationship to Dry Eye Clinical Severity.

The purpose of this study is to compare visual versus software detection of non-invasive tear break-up with the KOWA DR-1α tear interferometer and investigate the relationship between non-invasive tear break-up time (NIBUT) and dry eye clinical severity. Tear interferometry with the KOWA DR-1α, together with a standardized comprehensive ocular surface/tear evaluation, was performed on 348 consecutive eyes. Investigator visually detected or software detected non-invasive tear break-up and NIBUT were measured and compared on a subset of these examinations. The relationship between software-detected NIBUT and categorical dry eye severity based on irritation symptoms and corneal and conjunctival dye staining scores was determined. The sensitivity of visual (frame-by-frame) or software detected non-invasive tear break-up in eyes with tear instability (FBUT < 10) was similar (range 63−69%). NIBUT, measured visually or by software, had a correlation coefficient of 0.87. NIBUT was significantly lower in severity levels 2 and 3 compared to levels 0 + 1, and level 3 was significantly lower than level 2. In conclusion, there is a good correlation between investigator visually detected and software-detected tear break-up and tear break-up time in the KOWA DR-1α interferometric fringe images. Software-detected NIBUT is a clinically relevant measure of dry eye clinical severity.