Table 2.
Accuracy of CNN Models to Segment Retinal Layers Compared to Human Graders
| Central 6 mm | Full B-Scan | ||||
|---|---|---|---|---|---|
| of B-Scan, | Width (9 mm), | ||||
| % Accuracy | % Accuracy | ||||
| Characteristic | Mean | SD | Mean | SD | |
| U-Net | ILM-dINL | 96.0 | 4.0 | 94.8 | 4.1 |
| dINL-EZ | 93.5 | 5.4 | 93.1 | 5.1 | |
| OS (EZ-pPRE) | 85.9 | 13.6 | 85.4 | 13.6 | |
| RPE (pRPE-BM) | 87.7 | 4.5 | 86.8 | 4.7 | |
| SW | ILM-dINL | 94.8 | 7.1 | 92.8 | 7.7 |
| dINL-EZ | 93.3 | 6.8 | 91.8 | 7.0 | |
| OS (EZ-pPRE) | 88.5 | 9.0 | 87.9 | 9.6 | |
| RPE (pRPE-BM) | 86.2 | 6.4 | 83.5 | 7.1 | |
| Hybrid model | ILM-dINL | 97.0 | 1.1 | 95.7 | 3.2 |
| dINL-EZ | 94.1 | 5.3 | 93.7 | 5.4 | |
| OS (EZ-pPRE) | 87.0 | 10.3 | 86.4 | 10.4 | |
| RPE (pRPE-BM) | 87.9 | 4.5 | 86.7 | 5.4 | |
Mean accuracy was obtained by averaging all individual accuracy across all 160 test B-scan images and two graders.