Related Experiment Videos
A Deep Learning Approach in Rebubbling After Descemet's Membrane Endothelial Keratoplasty
Takahiko Hayashi1, Hitoshi Tabuchi, Hiroki Masumoto
1Department of Ophthalmology (T.H., N.K.), Yokohama Minami Kyosai Hospital, Kanagawa, Japan; Department of Ophthalmology (T.H., S.I., H. Takahashi), Jichi Medical University, Tochigi, Japan; Department of Ophthalmology (T.H.), Yokohama City University, Kanagawa, Japan; Department of Ophthalmology (H. Tabuchi, H.M.), Tsukazaki Hospital, Himeji, Japan; Graduate School of Engineering (S.M.), University of Hyogo; Department of Ophthalmology (I.O.), Heart Life Hospital, Okinawa, Japan; and Minami Aoyama Eye Clinic (N.K.), Tokyo, Japan.
Purpose:
To evaluate the efficacy of deep learning in judging the need for rebubbling after Descemet's endothelial membrane keratoplasty (DMEK).
Methods:
This retrospective study included eyes that underwent rebubbling after DMEK (rebubbling group: RB group) and the same number of eyes that did not require rebubbling (non-RB group), based on medical records. To classify the RB group, randomly selected images from anterior segment optical coherence tomography at postoperative day 5 were evaluated by corneal specialists. The criterion for rebubbling was the condition where graft detachment reached the central 4.0-mm pupil area. We trained nine types of deep neural network structures (VGG16, VGG19, ResNet50, InceptionV3, InceptionResNetV2, Xception, DenseNet121, DenseNet169, and DenseNet201) and built nine models. Using each model, we tested the validation data and evaluated the model.
Results:
This study included 496 images (31 eyes from 24 patients) in the RB group and 496 images (31 eyes from 29 patients) in the non-RB group. Because 16 picture images were obtained from the same point of each eye, a total of 992 images were obtained. The VGG19 model was found to have the highest area under the receiver operating characteristic curve (AUC) of all models. The AUC, sensitivity, and specificity of the VGG19 model were 0.964, 0.967, and 0.915, respectively, whereas those of the best ensemble model were 0.956, 0.913, and 0.921, respectively.
Conclusions:
This automated system that enables the physician to be aware of the requirement of RB might be clinically useful.