Can a Deep Learning Algorithm Quantify Glaucoma Damage?
A novel computer approach was able to use spectral-domain optical coherence tomography (SDOCT) data to train a deep learning algorithm to quantify glaucomatous structural damage on optic disc photographs in a cross-sectional study involving nearly 1,200 individuals.
Paired optic disc photos and SDOCT retinal nerve fiber layer (RNFL) scans were allocated into training (80%) and test (20%) sets. A deep learning algorithm was trained to assess optic disc photographs and predict SDOCT average RNFL thickness. The ability of the algorithm to differentiate between eyes with glaucomatous visual field loss from healthy eyes was assessed. Among the results:
- There was no difference between the mean prediction of average RNFL thickness from optic disc photos in the test and training sets (83.3 and 82.5 μm, respectively).
- Predicted and observed RNFL thickness values were highly correlated, with mean absolute error in the prediction set of 7.39 μm.
- The areas under the receiver operating characteristic curves for discriminating glaucomatous from healthy eyes with the computer algorithm predictions and actual SDOCT average RNFL thickness measurements were similar (0.944 and 0.940, respectively).
The authors concluded that the approach overcomes the need for human labeling of ocular samples and could be useful in other areas of ophthalmology.
Medeiros FA, Jammal AA, Thompson AC. From Machine to Machine: An OCT-trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs. Ophthalmology. [Published online ahead of print December 20, 2018] doi: 10.1016/j.ophtha.2018.12.033.