TABLE 3.

Summary of Detection and Segmentation Metrics (Mean Value ± Standard Deviation)

A: Voxel-by-voxel detection accuracy using ROC statistics*
# of metastases	AUC	Sensitivity	Specificity
1 to 3	0.99 ± 0.01	98 ± 3%	98 ± 2%
4 to 10	0.97 ± 0.05	92 ± 10%	97 ± 3%
>10	0.97 ± 0.03	92 ± 7%	95 ± 3%
All cases	0.98 ± 0.04	94 ± 8%	97 ± 3%
B: Detection and segmentation accuracy at an optimal probability threshold**
# of metastases	Dice	Recall	Precision
1 to 3	0.76 ± 0.20	0.54 ± .026	0.79 ± 0.27
4 to 10	0.83 ± 0.04	0.59 ± 0.21	0.76 ± 0.22
>10	0.78 ± 0.05	0.44 ± 0.18	0.81 ± 0.11
All cases	0.79 ± 0.12	0.53 ± 0.22	0.79 ± 0.20
C: Lesion-by-lesion detection accuracy at an optimal probability threshold**
# of metastases	Sensitivity	FP (no size limit)	FP (10 mm3 size limit)
1 to 3	92 ± 25%	3.2 ± 4.0	1.7 ± 2.0
4 to 10	81 ± 19%	8.5 ± 9.8	4.4 ± 6.0
>10	76 ± 20%	13.1 ± 18.9	4.1 ± 10.3
All cases	83 ± 22%	8.3 ± 12.9	3.4 ± 7.0

AUC = area under the receiver operating characteristic (ROC) curve; FP = false positive.

^*Sensitivity and specificity were determined by using the maximum value of Youden’s index.

^**The metrics were estimated using an optimal probability threshold of 0.93, as determined from the development set.