Table 4.
Models | Acc. | AUC | Sens. | Spec. | Prec. | F | MCC | DOR | Redux. |
---|---|---|---|---|---|---|---|---|---|
VGG-16-Baseline | 0.8828 | 0.9434 | 0.8766 | 0.8919 | 0.9221 | 0.8988 | 0.7612 (0.7106, 0.8118) |
58.575 | |
VGG-16-Fine-tuned | 0.9231 | 0.9532 | 0.9692 | 0.8559 | 0.9076 | 0.9374 |
0.8411
(0.7977, 0.8845) |
186.4375 | 0 |
VGG-19-Baseline | 0.9011 | 0.9441 | 0.9383 | 0.8469 | 0.8995 | 0.9185 | 0.7942 (0.7462, 0.8422) |
84.0471 | |
VGG-19-Fine-tuned | 0.9158 | 0.963 | 0.9198 | 0.9100 | 0.9372 | 0.9284 | 0.8264 (0.7814, 0.8714) |
115.7616 | 0 |
Inception-V3-Baseline | 0.8754 | 0.9304 | 0.9198 | 0.8109 | 0.8765 | 0.8976 | 0.7404 (0.6883, 0.7925) |
49.1209 | |
Inception-V3-Fine-tuned | 0.9048 | 0.9456 | 0.9198 | 0.8829 | 0.9198 | 0.9198 | 0.8027 (0.7554, 0.8500) |
86.4024 | 42.82 |
DenseNet-121-Baseline | 0.8645 | 0.9288 | 0.8519 | 0.8829 | 0.914 | 0.8818 | 0.7260 (0.6730, 0.7790) |
43.3462 | |
DenseNet-121-Fine-tuned | 0.8974 | 0.9399 | 0.9383 | 0.8379 | 0.8942 | 0.9157 | 0.7866 (0.7379, 0.8353) |
78.5334 | 56.52 |
NasNet-Mobile-Baseline | 0.8828 | 0.9403 | 0.9013 | 0.8559 | 0.9013 | 0.9013 | 0.7571 (0.7062, 0.8080) |
54.1797 | |
NasNet-Mobile-Fine-tuned | 0.8865 | 0.9258 | 0.9692 | 0.7658 | 0.858 | 0.9102 | 0.7679 (0.7178, 0.8180) |
102.6539 | 11.64 |
ResNet-18-Baseline | 0.8865 | 0.9371 | 0.9075 | 0.8559 | 0.9019 | 0.9047 | 0.7644 (0.7140, 0.8148) |
58.1875 | |
ResNet-18-Fine-tuned | 0.9048 | 0.9416 | 0.9013 | 0.9100 | 0.9359 | 0.9183 | 0.8052 (0.7582, 0.8522) |
92.1625 | 44.56 |
MobileNet-V2-Baseline | 0.8645 | 0.9188 | 0.8766 | 0.8469 | 0.8931 | 0.8848 | 0.7206 (0.6673, 0.7739) |
39.2589 | |
MobileNet-V2-Fine-tuned | 0.8865 | 0.9242 | 0.9754 | 0.7568 | 0.8541 | 0.9107 | 0.7694 (0.7194, 0.8194) |
122.8889 | 36.35 |
EfficientNet-B0-Baseline | 0.9194 | 0.9548 | 0.9383 | 0.8919 | 0.9269 | 0.9326 | 0.8327 (0.7884, 0.8870) |
125.4 | |
EfficientNet-B0-Fine-tuned | 0.9194 | 0.9442 | 0.963 | 0.8559 | 0.907 | 0.9342 | 0.8331 (0.7888, 0.8774) |
154.375 | 44.17 |
Data in parenthesis are 95% CI for the MCC values measured as the binomial (Clopper–Pearson’s) “exact” method corresponding to separate 2-sided CI with individual coverage probabilities of √0.95. Redux. = reduction in trainable parameters in%. The baseline signifies fine-tuning out-of-the-box ImageNet-pretrained CNNs toward this classification task. The best performances are denoted by bold numerical values in the corresponding columns. Except for the NasNet-Mobile and EfficientNet-B0 models, all other fine-tuned models demonstrated statistically significantly superior performance (p < 0.05) for their MCC metric compared to their baseline counterparts.