Table 4.

Unsupervised domain adaptation for TCGA (w/o UP) → UP and TCGA → RCINJ using color normalization and adversarial adaptation.

	TCGA (w/o UP) → UP	TCGA → RCINJ
Baseline	54.3	56.3
Macenko (Macenko et al., 2009) 1-Ensemble	65.7 ± 11.9	51.3 ±6.1
Macenko (Macenko et al., 2009) 2-Ensemble	70.0 ± 5.9	53.8 ±8.5
Macenko (Macenko et al., 2009) 5-Ensemble	72.3 ± 3.8	55.0 ± 7.3
Macenko (Macenko et al., 2009) 10-Ensemble	72.6 ± 2.3	55.0 ± 4.7
SPCN (Vahadane et al., 2016) 1-Ensemble	70.0 ± 7.3	56.3 ± 13.4
SPCN (Vahadane et al., 2016) 2-Ensemble	71.7 ± 6.7	55.0 ± 15.3
SPCN (Vahadane et al., 2016) 5-Ensemble	72.9 ± 2.6	55.6 ± 9.8
SPCN (Vahadane et al., 2016) 10-Ensemble	73.4 ± 1.8	54.4 ± 8.4
Color augmentation (Liu et al., 2017)	74.5	56.3
Generate-to-Adapt (Sankaranarayanan et al., 2018)	71.7	62.5
${\mathcal{L}}_{\mathrm{\text{a}}}$ only	71.4± 1.1	62.5 ± 2.5
${\mathcal{L}}_{\mathrm{\text{t}}}$	77.1± 1.1	75.0 ± 2.5

The classification accuracy of two color normalization methods including Macenko (Macenko et al., 2009) and SPCN (Vahadane et al., 2016) with different number of ensembles, and the target network with adversarial loss (${\mathcal{L}}_{\mathrm{\text{a}}}$) only and the target network with adversarial loass and Siamese loss together (${\mathcal{L}}_{\mathrm{\text{t}}}$) are shown for two sets of adaptations. We also compare our approach with color augmentation (Liu et al., 2017). Our proposed approach has a better performance than other state-of-the-art study (Sankaranarayanan et al., 2018) on the unsupervised adaptation task.