Purchase this article with an account.
Hironobu Tampo, Hidenori Takahashi, Yasuo Yanagi, Shin-ichi Sakamoto, Satoru Inoda, Hidetoshi Kawashima, Yuji Inoue, Yusuke Arai, Ryota Takahashi, Megumi Soeta; Deep-learning estimation of choroidal thickness from color fundus photographs. Invest. Ophthalmol. Vis. Sci. 2017;58(8):685.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The accuracy of general image classification has drastically improved since the advent of automatic extraction by deep learning to identify characteristic points in images. In 2015, artificial intelligence by deep learning was reported to be more accurate than human beings for the first time. Few reports have considered the application of deep-learning analysis to fundus photographs. Here, we applied deep learning to automate estimation of choroidal thickness from color fundus photographs.
We used images from 597 subjects who underwent same-day optical coherence tomography examination (swept-source OCT) [DRI-OCT (TOPCON, TOKYO)] and color fundus photography using a fundus camera [VX-10 (KOWA, AICHI)] at the outpatient clinic of Jichi Medical University hospital. From these images, those showing any of age-related macular degeneration, central serous chorioretinopathy, and proliferative diabetes retinopathy were excluded. Central choroidal thickness (CCT) was measured manually with OCT. When more than 6 images were available of a single eye, 5 of those images were used for classifier training. Forty eyes had only 1 associated fundus photograph. Images of the 3 eyes with the highest variance in CCT (due to diurnal variance) were excluded. Of the fundus photographs, 546. were used for deep learning. The deep learning system was run on a workstation (graphical processing units: 4) for 51 h. The 51 excluded photographs were used for validation.
The correlation coefficient for the 40 single-photograph eyes was 0.69. The same-eye correlation coefficients for the 3 eyes with highest CCT variance were 0.98, -0.96, and 0.45, in order of decreasing variance.
Artificial intelligence by deep learning estimated CCT from the color fundus photographs with good accuracy. However, single-eye diurnal CCT variation was not estimated well.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only