The Application of a Deep Learning Algorithm for the Segmentation of Retinal Nerve Fiber Layer Across Different Optic Neuropathies.

PURPOSE: To determine the ability of a deep learning (DL) algorithm to segment retinal nerve fiber layer (RNFL) from optical coherence tomography (OCT) scans in glaucomatous optic neuropathies and anterior optic neuropathies with optic disc edema.

METHODS

RNFL-Net was developed after preprocessing and automatically removing blood vessels from peripapillary OCT B-scans. It was trained and validated on 1065 RNFL OCT B-scans, and its performance was assessed using 265 test scans. Two different datasets were used for external testing.

RESULTS

The study involved 106 eyes from healthy controls, 118 eyes with optic disc edema, and 60 eyes with glaucoma for training and validation. The segmentation method achieved a Dice coefficient of 0.95 for the validation dataset and 0.92 for test images when compared with manual segmentation. In measuring RNFL thickness in glaucoma-affected eyes, RNFL-Net showed a mean absolute error (MAE) of 6.21 µm and a mean absolute percentage error (MAPE) of 11.24%. The standard OCT device had MAE of 11.05 µm and MAPE of 16.8%. For optic disc edema, the RNFL-Net MAE was 13.04 µm and MAPE 5.71%, whereas the OCT device reported MAE of 22.94 µm and MAPE of 11.2%. For the external validation data, MAE values for glaucoma (n = 157) and disc edema (n = 32) cases were 7.19 ± 0.14 and 15.41 ± 0.32, respectively.

CONCLUSIONS

RNFL-Net can accurately segment RNFL, whereas standard OCT devices produce lower measurements, especially in disc edema.

TRANSLATIONAL RELEVANCE: RNFL thickness measurements from RNFL-Net matched the ground truth in glaucoma and optic disc edema cases.