Table 2.
The performance comparisons of the 3-D DLS using different training strategies in both primary and external validations.
| Strategies | AUC (95% CI) | Sensitivity, % (95% CI) | Specificity, % (95% CI) | Accuracy, % (95% CI) |
|---|---|---|---|---|
| ResNet with original volumes | ||||
| Primary validation | 0.640 (0.594 to 0.686) | 53.3% (46.7 to 60.0) | 67.5% (63.9 to 71.3) | 63.9% (60.9 to 66.9) |
| External validation 1 | 0.535 (0.484 to 0.587) | 45.3% (1.0 to 66.3) | 67.4% (45.9 to 98.1) | 62.2% (50.9 to 76.9) |
| External validation 2 | 0.697 (0.616 to 0.779) | 60.0% (45.0 to 80.0) | 80.7% (57.6 to 89.1) | 76.2% (61.4 to 82.2) |
| ResNet with denoised volumes | ||||
| Primary validation | 0.806 (0.772 to 0.841) | 71.9% (65.7 to 77.6) | 74.9% (71.6 to 78.2) | 74.1% (71.2 to 77.1) |
| External validation 1 | 0.645 (0.598 to 0.693) | 51.4% (41.4 to 62.4) | 74.9% (66.8 to 82.1) | 68.9% (63.6 to 73.1) |
| External validation 2 | 0.793 (0.732 to 0.853) | 75.0% (58.3 to 91.7) | 74.8% (53.8 to 87.4) | 75.2% (60.7 to 83.2) |
| SE-ResNet with original volumes | ||||
| Primary validation | 0.905 (0.875 to 0.934) | 82.9% (77.1 to 89.1) | 93.3% (88.2 to 96.2) | 90.6% (87.3 to 92.6) |
| External validation 1 | 0.815 (0.779 to 0.852) | 68.0% (59.7 to 80.7) | 83.2% (68.1 to 87.9) | 79.2% (70.6 to 82.3) |
| External validation 2 | 0.858 (0.801 to 0.915) | 78.3% (63.3 to 90.0) | 81.1% (73.5 to 92.9) | 81.4% (75.2 to 88.7) |
| SE-ResNet with denoised volumes | ||||
| Primary validation | 0.943 (0.925 to 0.961) | 83.1% (77.1 to 88.6) | 93.1% (87.6 to 96.4) | 90.6% (87.1 to 92.7) |
| External validation 1 | 0.811 (0.774 to 0.847) | 68.0% (59.1 to 80.1) | 83.2% (69.1 to 88.3) | 79.2% (71.3 to 82.7) |
| External validation 2 | 0.844 (0.784 to 0.847) | 78.3% (63.3 to 91.7) | 81.5% (73.5 to 93.7) | 81.5% (75.2 to 89.3) |
| SE-ResNeXt with original volumes | ||||
| Primary validation | 0.938 (0.919 to 0.957) | 84.8% (78.6 to 91.0) | 91.6% (86.2 to 95.6) | 89.7% (86.4 to 92.4) |
| External validation 1 | 0.801 (0.764 to 0.838) | 68.0% (53.6 to 86.2) | 78.9% (58.7 to 90.6) | 76.1% (66.1 to 82.4) |
| External validation 2 | 0.854 (0.796 to 0.912) | 76.7% (61.7 to 90.0) | 84.9% (74.0 to 95.4) | 83.2% (75.8 to 90.3) |
| SE-ResNeXt with denoised volumes | ||||
| Primary validation | 0.954 (0.938 to 0.970) | 86.2% (80.0 to 92.4) | 92.6% (86.8 to 96.9) | 91.0% (87.3 to 93.5) |
| External validation 1 | 0.816 (0.780 to 0.852) | 69.1% (58.0 to 84.0) | 81.3% (64.0 to 89.4) | 78.2% (68.8 to 82.7) |
| External validation 2 | 0.857 (0.800 to 0.914) | 78.3% (61.7 to 91.7) | 82.8% (71.9 to 94.6) | 82.6% (74.2 to 89.9) |
Note: AUC, area under the receiver operator characteristic curve; CI, confidence interval; DLS, deep learning system.
Primary validation: CUHK Eye Center and HKEH;
External validation 1: PWH, Hong Kong;
External validation 2: TMEC, Hong Kong.
The bold values were the highest values in respective categories.