Skip to main content
. 2020 Jul 3;14:171. doi: 10.3389/fncel.2020.00171

Figure 3.

Figure 3

Comparison of different convolutional neural network (CNN) architectures. (A) Loss curves and receiver operating characteristic-area under the curve (AUC) training curves for VGG19, ResNET50v2, DenseNet121, and Xception. (B) Comparison summary of three different CNNs using 10-fold cross-validation. The mean AUC scores were 0.93 ± 0.03 vs. 0.91 ± 0.04 vs. 0.92 ± 0.04 (P = 0.3) for ResNET50v2, DenseNet121, and Xception, respectively; the mean F1 scores were 0.89 ± 0.02 vs. 0.88 ± 0.04 vs. 0.88 ± 0.04 for ResNET50v2, DenseNet121, and Xception, respectively; the mean accuracy scores were 0.85 ± 0.03 vs. 0.83 ± 0.05 vs. 0.83 ± 0.06 for ResNET50v2, DenseNet121, and Xception, respectively; the mean Matthews correlation coefficients were 0.64 ± 0.08 vs. 0.62 ± 0.11 vs. 0.63 ± 0.12 for ResNET50v2, DenseNet121, and Xception, respectively. Each dot on the graph corresponds to one cross-validation step. ns, not significant (P-value > 0.05 on Friedman statistical test).