Table 3.

Per-image AUCs of whole image classifiers using the VGG16 as patch classifiers in the independent test set.

Patch set	Block 1	Block 2	AUC [95% CI]	A-AUC [95% CI]	#Epochs
S10	512 × 3	512 × 3	0.81 [0.77, 0.84]	0.82 [0.78, 0.85]	91
S10	256 × 1	128 × 1	0.85 [0.81, 0.88]	0.86 [0.83, 0.89]	61
S10	128 × 1	64 × 1	0.84 [0.80, 0.87]	0.86 [0.82, 0.89]	142
S10	[512-512-1024] × 2	[512-512-1024] × 2	0.85 [0.82, 0.88]	0.88, [0.85, 0.91]	165
Add heatmap and fully connected (FC) layers on top (S10 patch set)
Pool size	FC1	FC2
5 × 5	64	32	0.71 [0.66, 0.75]	NA	26
2 × 2	512	256	0.68 [0.63, 0.73]	NA	27
1 × 1	2048	1024	0.70 [0.65, 0.74]	NA	50

#Epochs indicates the epoch when the highest AUC was reached in the validation set. The best performing models are shown in boldface.