Abstract
Purpose :
We previously developed a deep learning algorithm using data from a single institution to estimate choroidal thickness from color fundus photographs (CFPs). This study aimed to validate the algorithm using data from multiple institutions.
Methods :
The algorithm was trained using 1000 CFPs and corresponding choroidal thickness measurements by optical coherence tomography (OCT) where taken on the same day at Jichi Medical University (institution A; fundus camera, VX-10, Nagoya, Japan). Exclusion criteria were any signs of subretinal hemorrhage, central serous detachment, retinal pigment epithelial detachment, and macular edema. Most of CFPs were normal fellow eyes of the outpatients. All other CFPs of good quality were included. Eyes with moderate retinal hemorrhage, drusen, moderate exudates, and glaucoma were included. Validation study was performed using data from Kagoshima University (B; TRC-50LX and TRC-NW7S, TOPCON, Tokyo, Japan), Saitama Medical Center (C; VX-10), and Ibaraki Seinan Medical Center (D; 3D OCT-2000, TOPCON). One hundred thirty five, 148, and 129 CFPs were used from institution B, C and D, respectively. Absolute error (difference between predicted and actual choroidal thickness) was calculated and standard deviation (SD) was compared using F test with the data from institution A as a control. Baseline characteristics of the subjects were compared using t-test for continuous values such as choroidal thickness, axial length, and chi-square test for gender.
Results :
The SDs of absolute error at institution B, C, and D were 71, 83, and 128 µm, which were significantly larger than the SD of 21 µm at institution A (P < 0.001). At institution A, B, C, and D, respectively, mean choroidal thickness (SD, P value compared with A) was 241 (99), 306 (74, P < 0.001), 245 (104), and 339 (118, P < 0.001) µm; mean axial length was 24 (1.9), 25 (1.4, P < 0.001), 24 (1.2), and 24 (1.1) mm; mean age was 65 (15), 26 (3.8, P < 0.001), 56 (18, P < 0.001), and 69 (15, P = 0.02) years; and male sex ratio was 61%, 66%, 51% (P = 0.01), and 51% (P = 0.05).
Conclusions :
This study demonstrates that accuracy of estimation of choroidal thickness was worse when CFP was taken under different condition or from the patients with different attributes, highlightening the importance of dataset selection for the developemt of deep-learning based algorism.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.