To develop a deep learning method to predict visual field (VF) from wide-angle swept-source optical coherence tomography (SS-OCT) and compare the performance of three Google Inception architectures.
Methods
Three deep learning models (with Inception-ResNet-v2, Inception-v3, and Inception-v4) were trained to predict 24-2 VF from the macular ganglion cell-inner plexiform layer and the peripapillary retinal nerve fibre layer map obtained by SS-OCT. The prediction performance of the three models was evaluated by using the root mean square error (RMSE) between the actual and predicted VF. The performance was also compared among different glaucoma severities and Garway-Heath sectorizations.
Results
The training dataset comprised images of 2220 eyes from 1120 subjects, and the test dataset was obtained from another 305 subjects (305 eyes). In all subjects, the global prediction errors (RMSEs) were 4.44 ± 2.09 dB, 4.78 ± 2.38 dB, and 4.85 ± 2.66 dB for the Inception-ResNet-v2, Inception-v3, and Inception-v4 architectures, respectively, and the prediction error of Inception-ResNet-v2 was significantly lower than the other two (P < 0.001). As glaucoma progressed, the prediction error of all three architectures significantly worsened to 6.59 dB, 7.33 dB, and 7.79 dB, respectively. In the analysis of sectors, the nasal sector had the lowest prediction error, followed by the superotemporal sector.
Conclusions
Inception-ResNet-v2 achieved the best performance, and the global prediction error (RMSE) was 4.44 dB. As glaucoma progressed, the prediction error became larger. This method may help clinicians determine VF, particularly for patients who are unable to undergo a physical VF test.
