Table 5.

Comparing the prediction power of the employed methods.

Prediction Performance Comparison (AUROC)
Fine-tuning	Context-info	Best conventional radiomics	Best end-to-end deep learning model	Best deep-feature-based radiomics	Best hybrid model
Before	No	0.792 ± 0.025	0.801 [0.777,0.824]	0.753 [0.743,0.775]	0.817 ± 0.032
After	No	0.911 ± 0.016	–	0.906 [0.890,0.921]	0.914 ± 0.015
Before	Yes	0.777 ± 0.017	0.806 [0.788,0.827]	0.761 [0.736,0.779]	0.780 ± 0.022
After	Yes	0.916 ± 0.011	0.824 [0.798,0.837]	0.927 [0.912,0.940]	0.929 ± 0.013

Note that the “best end-to-end deep learning model” column presents the performance of two single pathway models trained with target and context nodule images separately and one dual pathway model trained with both target and context images simultaneously. Lower and upper limits of confidence interval at 95% level are indicated in square brackets.