Sachiko Maruoka1, Hitoshi Tabuchi1,2, Daisuke Nagasato1,2, Hiroki Masumoto1,2, Taiichiro Chikama1,3, Akiko Kawai1, Naoko Oishi1, Toshi Maruyama1,4, Yoshitake Kato1, Takahiko Hayashi1,2,5, Chikako Katakami1. 1. Department of Ophthalmology, Tsukazaki Hospital, Himeji, Japan. 2. Department of Technology and Design Thinking for Medicine, Hiroshima University Graduate School, Hiroshima, Japan. 3. Department of Ophthalmology and Visual Sciences, Graduate School of Biomedical Sciences, Hiroshima University, Hiroshima, Japan. 4. Department of Ophthalmology, Maruyama Eye Clinic, Maebashi, Japan; and. 5. Department of Ophthalmology, Yokohama Minami Kyosai Hospital, Kanagawa, Japan.
Abstract
PURPOSE: To evaluate the ability of deep learning (DL) models to detect obstructive meibomian gland dysfunction (MGD) using in vivo laser confocal microscopy images. METHODS: For this study, we included 137 images from 137 individuals with obstructive MGD (mean age, 49.9 ± 17.7 years; 44 men and 93 women) and 84 images from 84 individuals with normal meibomian glands (mean age, 53.3 ± 19.6 years; 29 men and 55 women). We constructed and trained 9 different network structures and used single and ensemble DL models and calculated the area under the curve, sensitivity, and specificity to compare the diagnostic abilities of the DL. RESULTS: For the single DL model (the highest model; DenseNet-201), the area under the curve, sensitivity, and specificity for diagnosing obstructive MGD were 0.966%, 94.2%, and 82.1%, respectively, and for the ensemble DL model (the highest ensemble model; VGG16, DenseNet-169, DenseNet-201, and InceptionV3), 0.981%, 92.1%, and 98.8%, respectively. CONCLUSIONS: Our network combining DL and in vivo laser confocal microscopy learned to differentiate between images of healthy meibomian glands and images of obstructive MGD with a high level of accuracy that may allow for automatic obstructive MGD diagnoses in patients in the future.
PURPOSE: To evaluate the ability of deep learning (DL) models to detect obstructive meibomian gland dysfunction (MGD) using in vivo laser confocal microscopy images. METHODS: For this study, we included 137 images from 137 individuals with obstructive MGD (mean age, 49.9 ± 17.7 years; 44 men and 93 women) and 84 images from 84 individuals with normal meibomian glands (mean age, 53.3 ± 19.6 years; 29 men and 55 women). We constructed and trained 9 different network structures and used single and ensemble DL models and calculated the area under the curve, sensitivity, and specificity to compare the diagnostic abilities of the DL. RESULTS: For the single DL model (the highest model; DenseNet-201), the area under the curve, sensitivity, and specificity for diagnosing obstructive MGD were 0.966%, 94.2%, and 82.1%, respectively, and for the ensemble DL model (the highest ensemble model; VGG16, DenseNet-169, DenseNet-201, and InceptionV3), 0.981%, 92.1%, and 98.8%, respectively. CONCLUSIONS: Our network combining DL and in vivo laser confocal microscopy learned to differentiate between images of healthy meibomian glands and images of obstructive MGD with a high level of accuracy that may allow for automatic obstructive MGD diagnoses in patients in the future.