Cascaded-Recalibrated Multiple Instance Deep Model for Pathologic-Level Lung Cancer Prediction in CT Images

Table 4

Performance of cascaded-recalibrated MIL model based on the top- attribute sources transfer.

Top-	Sources Target(cascaded-recalibrated MIL)	Accuracy	AUC	F1-score	Runtime
					Training	Test (s)
Top-1	Tex ⟶ Pmal	0.846 ± 0.037	0.849 ± 0.027	0.813 ± 0.037	6 min 31 s	0.0035
Top-2	Tex + Sph ⟶ Pmal	0.863 ± 0.051	0.867 ± 0.070	0.834 ± 0.062	9 min 4 s	0.0046
Top-3	Tex + Sph + Mal ⟶ Pmal	0.880 ± 0.032	0.877 ± 0.036	0.849 ± 0.037	12 min 0 s	0.0066
Top-4	Tex + Sph + Mal + Lob ⟶ Pmal	0.847 ± 0.056	0.842 ± 0.053	0.817 ± 0.066	15 min 17 s	0.0083
Top-5	Tex + Sph + Mal + Lob + Spi ⟶ Pmal	0.846 ± 0.020	0.82 ± 0.0460	0.800 ± 0.031	18 min 27 s	0.0098
Top-6	Tex + Sph + Mal + Lob + Spi + Mar ⟶ Pmal	0.837 ± 0.065	0.833 ± 0.065	0.808 ± 0.071	21 min 42 s	0.0103
Top-7	Tex + Sph + Mal + Lob + Spi + Mar + Cal ⟶ Pmal	0.829 ± 0.028	0.821 ± 0.056	0.791 ± 0.036	24 min 30 s	0.0133
Top-8	Tex + Sph + Mal + Lob + Spi + Mar + Cal + Sub ⟶ Pmal	0.828 ± 0.041	0.821 ± 0.043	0.798 ± 0.047	28 min 16 s	0.0146
Top-9	Tex + Sph + Mal + Lob + Spi + Mar + Cal + Sub + Int ⟶ Pmal	0.803 ± 0.020	0.816 ± 0.013	0.770 ± 0.028	31 min 27 s	0.0161

The bold values mean the superiority to others.