Table 5. Performance evaluation of DL models for dyslexia neural-biomarker classification without/with smoothen and normalized dataset (mean ± SD after 10 repeated 10-fold CV).
| DL Model | Iteration (Epoch) | Accuracy (%) | Sensitivity (%) | Specificity (%) | F-Score (%) | Feature extraction time (mins) | |
|---|---|---|---|---|---|---|---|
| Inception-V3 | Without smoothing and MHN | 402 | 86.23±1.99 | 88.91±3.78 | 85.68±3.20 | 87.27±3.47 | 41.66 |
| With smoothing and MHN | 380 | 89.08±1.22 | 90.22±2.61 | 92.86±2.14 | 91.52±2.35 | 33.45* | |
| Cascaded CNN | Without smoothing and MHN | 498 | 80.91±2.31 | 91.74±1.88 | 93.11±3.04 | 92.42±2.23 | 44.02 |
| With smoothing and MHN | 550 | 91.21±0.89 | 93.11±2.64 | 92.95±2.46 | 93.03±2.55 | 31.78* | |
| ResNet50 | Without smoothing and MHN | 369 | 93.33±1.02 | 95.11±2.87 | 91.42±0.83 | 93.23±1.29 | 23.67 |
| With smoothing and MHN | 450 | 94.67±0.69* | 95.79±2.18* | 94.91±2.16* | 95.35±2.17* | 12.65* |
*Statistically significantly larger than the other two (p-value<0.05); 95% CI level. MHN-modified histogram normalization; CNN-convolutional neural network.