%0 Journal Article %A Akiba, Masahiro %A An, Guangzhou %A yokota, Hideo %A Omodaka, Kazuko %A Hashimoto, Kazuki %A Tsuda, Satoru %A Shiga, Yukihiro %A takada, naoko %A kikawa, tsutomu %A Nakazawa, Toru %T Evaluation of glaucoma diagnosis machine learning models based on color optical coherence tomography and color fundus images %B Investigative Ophthalmology & Visual Science %D 2019 %J Investigative Ophthalmology & Visual Science %V 60 %N 9 %P 1298-1298 %@ 1552-5783 %X To develop and evaluate deep learning based glaucoma diagnosis models, based on optical coherence tomography (OCT) data and color fundus images, aiming for clinical use. This study included 149 healthy (MD: -0.21±1.15 dB) and 208 open-angle glaucoma (OAG) (MD: -3.90±3.80 dB) eyes. All these eyes were captured with spectral-domain OCT (3D OCT-2000, Topcon), in protocol of macular and disc volumetric scans, as well as color fundus images. With the commercial OCT segmentation software, we created retinal nerve fiber layer (RNFL) thickness map from disc volumetric OCT data, and ganglion cell complex (GCC) thicknees map from macular volumetric OCT data. First, with the created images we transfer learned 3 separate convolutional neural network (CNN) models with well-known architecure of VGG19, which were pre-trained on ImageNet dataset. With these VGG19 models as feature extractors, images features were obtained in the top level of VGG19s for each kind of images, and concatenated for training a random forest (RF) to combine separate models. The area under receiver operating characteristic curve (AUC) of a 10-fold cross-validation (CV) was applied to evaluate our models. The 10-fold AUCs of the VGG19 models were 0.940 for color fundus images, 0.942 for cpRNFL thickness map, 0.944 for macular GCC thickness maps. The RF using all kinds of images achieved the highest 10-fold CV AUC of 0.961. We found by adjusting the threshold, our model got an excellent sensitivity for detecting glaucoma, with acceptable false positive ratio. Our proposed approach was efficient for automatic classification of healthy and glaucoma eyes based on OCT and color fundus images, and it might be used in assisting glaucoma diagnosis. The optimized threshold for our model in clinical use should be determined for our future work. This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019. %[ 4/12/2021