OCT-trained deep learning algorithm helpful for assessing glaucomatous neural damage

This study describes a new approach that uses data from spectral-domain (SD) OCT to train a deep-learning (DL) algorithm for quantifying glaucomatous structural damage.

Study design

The authors trained a DL neural network on 32,820 pairs of disc photos and RNFL scans to predict SD-OCT average RNFL thickness. The sample was divided into a validation plus training set and a test set. The DL performance was assessed in the test sample by evaluating correlation and agreement between the predicted and actual OCT measurements.

Outcomes

There was a strong correlation between predicted and observed mean average RNFL thickness (r=0.832), with a low mean absolute error of prediction (7.39 microns). After training on disc photographs, the DL determined normal vs. abnormal RNFL thickness with 83.7% accuracy.

Limitations

Disc photo quality was not assessed or controlled for. Additionally, the algorithm was trained to identify average RNFL thickness; thus, segmental loss in the face of normal average thickness could potentially misclassify glaucomatous nerves.

Clinical significance

This is an innovative approach that has the advantage of removing the subjectivity and poor reproducibility of human grading. It has potentially wide applications for glaucoma screening and assessment of optic nerve changes over time in practices that do not have access to SD-OCT.