Skip to main content
. Author manuscript; available in PMC: 2020 Jul 1.
Published in final edited form as: Comput Med Imaging Graph. 2019 May 28;75:66–73. doi: 10.1016/j.compmedimag.2019.05.005

Table I.

Performance summary (AUC scores) of existing approaches and our proposed SDFN using the NIH benchmark split

Wang et al. [11] Yao et al.[14] Gündel et al.[16] Entire CXR Lung region Proposed SDFN
Atelectasis 0.700 0.733 0.767 0.762 0.773 0.781
Cardiomegaly 0.814 0.856 0.883 0.878 0.868 0.885
Effusion 0.736 0.806 0.828 0.822 0.827 0.832
Infiltration 0.613 0.673 0.709 0.693 0.670 0.700
Mass 0.693 0.777 0.821 0.791 0.807 0.815
Nodule 0.669 0.724 0.758 0.744 0.759 0.765
Pneumonia 0.658 0.684 0.731 0.707 0.708 0.719
Pneumothorax 0.799 0.805 0.846 0.855 0.851 0.866
Consolidation 0.703 0.711 0.745 0.737 0.738 0.743
Edema 0.805 0.806 0.835 0.837 0.831 0.842
Emphysema 0.833 0.842 0.895 0.912 0.909 0.921
Fibrosis 0.786 0.743 0.818 0.826 0.825 0.835
Pleural Thickening 0.684 0.724 0.761 0.760 0.783 0.791
Hernia 0.872 0.775 0.896 0.902 0.912 0.911
Mean 0.740 0.761 0.807 0.802 0.804 0.815
*

Entire CXR and Lung region represent the fine-tuned DenseNets trained on the entire CXR images and local lung region images respectively.