Purchase this article with an account.
Samuel Johnson, Mohammad Shafkat Islam, Jui-Kai Wang, Thurtell Matthew, Randy H Kardon, Mona Garvin; Deep-Learning-Based Estimation of Regional Volumetric Information from 2D Fundus Photography in Cases of Optic Disc Swelling. Invest. Ophthalmol. Vis. Sci. 2019;60(9):3597.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
With the introduction of optical coherence tomography (OCT) as well as retinal-layer-segmentation algorithms, quantitative assessment for optic disc swelling has improved. However, OCT is not always available as its use is usually limited to specialized clinics. Previously, we published an approach using a random forest to estimate the retinal thickness between the inner limiting membrane (ILM) and retinal pigment epithelium (RPE) in fundus photographs (Johnson et al., OMIA, 2018). We now propose a new model using deep-learning techniques to predict thickness between the surfaces which avoids selection of key image features for model training.
A U-Net convolutional neural network (Ronneberger et al., MICCAI 2015) was trained using 78 (70 training, 8 validation) cropped fundus photographs (input, Fig. 1, top) and matching OCT-derived thickness maps (reference standard, Fig. 1, bottom) from recruited subjects with optic disc swelling at the University of Iowa Neuro-Ophthalmology Clinic. Images were preprocessed by using contrast limited adaptive histogram equalization as well as normalization. Hyperparameters were as follows: 500 epochs, 10e-3 learning rate, and a batch size of 8. The trained model was then used to predict OCT-based thickness maps on ten strictly withheld test images.
Model predictions on the test images were compared with OCT-derived volumetric measures: total retinal volume as well as regional volumes were calculated. For total retinal volume, a root-mean-square-error (RMSE) of 2.07 mm3 was achieved. When comparing regional volumes, the nasal, temporal, inferior, superior, and peripapillary regions had RMSEs of 0.75 mm3, 0.82 mm3, 0.85 mm3, 0.91 mm3, and 1.62 mm3 respectively.
The proposed deep-neural-network is capable of estimating retinal thickness at the pixel level in fundus photographs without using expert-designed image features. Since OCT is not available everywhere yet, this could be an alternative approach to gain insight for optic-disc-swelling assessment.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Fig 1: Top: Input 200x200 ONH region fundus photographs. Middle: The predicted thickness map for the above fundus image. Bottom: OCT reference standard thickness map. All volumes shown are in mm3. PRV (partial retinal volume) refers to the cumulative volume of all 4 quadrants. TRV is defined in the map image.
Fig 2: Workflow depiction with network architecture
This PDF is available to Subscribers Only