Table 2:

Stability and performance of machine learning optimizing repeatability using average of feature values extracted from prostate DWI with four feature ranking methods. The 95% confidence intervals are shown in parenthesis. Rows 1–4: Agreement between classifications and performance for Gleason Grade Group (GGG) 1 vs >1 classification. Rows 5–8: Agreement between classifications and performance for GGG 1,2 vs >2.

Feature ranking method	Pruned features	% of same labels	AUC (95% CI) Scan1	AUC (95% CI) Scan 2	AUC (95% CI) Scan 1 and 2
Wilcoxon rank-sum test	ADCk & K ICC>0.8 & Top 10	60.8	0.780 (0.607..0.952)	0.757 (0.572..0.942)	0.765 (0.640..0.890)
MRMR	ADCk & K ICC>0.8 & Top 10	80.4	0.781 (0.601..0.960)	0.725 (0.560..0.890)	0.747 (0.629..0.866)
Spearman ρ	ADCk & K ICC>0.8 & Top 10	72.5	0.686 (0.503..0.870)	0.723 (0.566..0.880)	0.706 (0.589..0.822)
AUC	ADCk & K ICC>0.8 & Top 10	64.7	0.782 (0.601..0.962)	0.709 (0.548..0.870)	0.743 (0.623..0.863)
Wilcoxon rank-sum test	ADCk & K ICC>0.8 & Top 10	72.5	0.705 (0.530..0.879)	0.791 (0.600..0.981)	0.745 (0.620..0.870)
MRMR	ADCk & K ICC>0.8 & Top 10	66.7	0.734 (0.550..0.918)	0.745 (0.557..0.934)	0.732 (0.606..0.858)
Spearman ρ	ADCk & K ICC>0.8 & Top 10	62.7	0.745 (0.551..0.940)	0.734 (0.563..0.905)	0.738 (0.613..0.863)
AUC	ADCk & K ICC>0.8 & Top 10	62.7	0.745 (0.551..0.940)	0.734 (0.563..0.905)	0.738 (0.613..0.863)