Purchase this article with an account.
Jonathan D Oakley, Daniel B Russakoff; Automated choroid segmentation of SD-OCT volumes using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB0109.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To automatically segment the choroid’s posterior boundary in SD-OCT and evaluate for accuracy across different retinal pathologies.
The posterior boundary of the choroid was manually delineated in 50 OCT volumes of the macula using Orion’s layer editing tool (Voxeleron, Pleasanton, CA). The volumes were chosen from a control population without retinal pathology with the following exclusions: axial motion issues, choroid boundary out of the field of view / not clearly defined.The OCT data was denoised and down-sampled to a volume of 128x128x128. Various deep convolutional neural network (CNN) U-Net architectures were explored/optimized to assign three labels to each OCT voxel: background, retina and choroid. 5-fold cross validation, where no eye from the same subject was in both training and test sets, was used to assess performance. Overall segmentation accuracy was reported using a dice coefficient to gauge the entire volume of overlap between the manual and automatic segmentations.To evaluate the method’s ability to generalize to unseen data and pathologies, the best architecture/model, trained on the previous 50 cases, was applied to a new data set comprising 30 volumes of mixed retinal pathologies acquired from a separate center.
During cross validation, the best segmentation network had average dice scores of 0.99 (STD=0.004), 0.99 (0.004) and 0.95 (0.019) for background, retina and choroid, respectively. On the unseen retinal pathology data, the dices scores were: 0.98 (0.009), 0.99 (0.006) and 0.91 (0.033). Example segmentations from the original data and the retinal pathology data are shown in Figs 1 and 2, respectively.
Quantification of choroidal thickness can inform upon a variety of chorioretinal diseases. Such assessment is challenging, particularly in SD-OCT devices, but with suitable pre-processing, deep learning offers a ready solution. Comparison to ground truth shows excellent correlation and generalizability, paving the way for wider use in OCT, enhanced depth imaging, and swept source devices.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
Fig (1): (a) automated retinal segmentation from the control group (N=50) of the ILM (red) and RPE (blue) (Orion) and manual choroid segmentation (green); (b) the posterior boundary of the choroid is replaced with the deep-learning result.
Fig (2): Example automated segmentation results based on the second data set of mixed retinal pathologies (N=30).
This PDF is available to Subscribers Only