Table 1 The average performance metrics with 95% confidence intervals (CIs) for the stepwise transfer learning and other methods on the fivefold cross-validation for predicting cases of refractory central serous chorioretinopathy (CSC).
AUC (95% CI) | Accuracy (%) (95% CI) | Sensitivity (%) (95% CI) | Specificity (%) (95% CI) | P value* | |
|---|---|---|---|---|---|
Stepwise transfer learning | |||||
ResNet50 | 0.839 (0.770‒0.895) | 83.0 (75.9‒88.7) | 67.7 (48.6‒83.3) | 87.1 (79.6‒92.6) | Ref. |
VGG16 | 0.827 (0.756‒0.884) | 83.7 (76.7‒89.3) | 64.5 (45.4‒80.8) | 88.8 (81.6‒93.9) | 0.539 |
Traditional transfer learning | |||||
ResNet50 | 0.808 (0.735‒0.868) | 81.0 (73.7‒87.0) | 64.5 (45.4‒80.8) | 85.3 (77.6‒91.2) | 0.223 |
VGG16 | 0.800 (0.726‒0.861) | 80.3 (72.9‒86.4) | 61.3 (42.2‒78.2) | 85.3 (77.6‒91.2) | 0.198 |
Xception | 0.816 (0.744‒0.875) | 81.6 (74.4‒87.5) | 64.5 (45.4‒80.8) | 86.2 (78.6‒91.9) | 0.337 |
EfficientNet-b0 | 0.796 (0.722‒0.858) | 82.3 (75.2‒88.1) | 58.1 (39.1‒75.5) | 88.8 (81.6‒93.9) | 0.149 |
Other architecture | |||||
NASNet-Large | 0.582 (0.498‒0.663) | 44.2 (36.0‒52.6) | 87.1 (70.2‒96.4) | 32.8 (24.3‒42.1) | 0.002 |
