Table 3.
CoAtNet-0-rw | EfficientNet-b5 | Baseline (ResNet50) | Baseline (ResNet50) | ||||
---|---|---|---|---|---|---|---|
Labels | LWBCE | Lours | LWBCE | Lours | LWBCE | Lours | LWBCE – Previous study [8]# |
AUROC | AUROC | AUROC | AUROC | AUROC | AUROC | AUROC | |
Atelectasis | 0.795 | 0.833 | 0.795 | 0.812 | 0.776 | 0.791 | 0.707 |
Cardiomegaly | 0.901 | 0.914 | 0.906 | 0.905 | 0.908 | 0.890 | 0.814 |
Consolidation | 0.787 | 0.809 | 0.798 | 0.802 | 0.786 | 0.788 | |
Edema | 0.903 | 0.912 | 0.905 | 0.908 | 0.898 | 0.899 | |
Effusion | 0.874 | 0.890 | 0.872 | 0.882 | 0.867 | 0.878 | 0.736 |
Emphysema | 0.927 | 0.940 | 0.906 | 0.912 | 0.872 | 0.871 | |
Fibrosis | 0.826 | 0.835 | 0.805 | 0.807 | 0.790 | 0.788 | |
Hernia | 0.865 | 0.791 | 0.826 | 0.837 | 0.847 | 0.768 | |
Infiltration | 0.686 | 0.715 | 0.699 | 0.712 | 0.688 | 0.705 | 0.613 |
Mass | 0.827 | 0.856 | 0.827 | 0.832 | 0.819 | 0.823 | 0.561 |
Nodule | 0.748 | 0.779 | 0.745 | 0.755 | 0.722 | 0.718 | 0.716 |
Pleural_Thickening | 0.804 | 0.819 | 0.810 | 0.810 | 0.789 | 0.790 | |
Pneumonia | 0.751 | 0.789 | 0.736 | 0.774 | 0.733 | 0.733 | 0.633 |
Pneumothorax | 0.870 | 0.902 | 0.878 | 0.886 | 0.857 | 0.859 | 0.789 |
Overall* | 0.826 ± 0.018 (0.786, 0.866) | 0.842 ± 0.017 (0.805, 0.879)a | 0.822 ± 0.018 (0.784, 0.860) | 0.831 ± 0.016 (0.796, 0.866)b | 0.811 ± 0.018 (0.772, 0.850) | 0.807 ± 0.017 (0.770, 0.844) | 0.696 ± 0.031 (0.623, 0.769)c |
Pearson correlation with the positive sample ratiod in the test set | −0.517 | −0.385 | −0.421 | −0.432 | −0.442 | −0.269 | −0.260 |
p value | 0.058 | 0.174 | 0.134 | 0.123 | 0.113 | 0.352 | 0.534 |
aThere are significant differences when compared with ResNet50 + LWBCE (p = 0.037), ResNet50 + Lours (p < 0.001), and results of ResNet50 + LWBCE in previous studies (p = 0.017)
bThere are significant differences when compared with ResNet50 + LWBCE (p = 0.004), ResNet50 + Lours (p = 0.004), and results of ResNet50 + LWBCE in previous studies (p = 0.028)
cThere are significant differences compared to the other six models (p ≤ 0.006)
dThe positive sample ratio was calculated by the count of positive samples over the total sample size. All statistical analyses were performed using repeated measures ANOVA tests with Bonferroni adjustments
#In this previous study, only 8 labels were involved
*Data were presented as mean ± standard error (95% confidence interval)