Table 3.

Quantitative evaluation of several unsupervised harmonization methods and an ablation of the proposed method using the held-out traveling subjects. Both SSIM and PSNR are computed on 3D volumes, and they are reported as “mean ”± “standard deviation ”.“median ” denotes the 2D-to-3D median fusing approach similar to Dewey et al. (2019). “fusion ” indicates our proposed 2D-to-3D fusing method described in Section 2.5. The proposed method shows significant improvements ($p<.05$, $N=10$) over all comparison methods based on paired Wilcoxon signed rank tests, with exceptions labeled by ^†. Bold numbers indicate the best mean performance.

	Combine	Perceptual		Site $C\to D$		Site $D\to C$		Site $E\to F$		Site $F\to E$
No har	–	–	SSIM	0.7647	±0.0151	0.7647	±0.0151	0.8517	±0.0343	0.8517	±0.0343†
			PSNR	25.62	±0.85	25.62	±0.85	27.86	±1.24	27.86	±1.24
Hist	–	–	SSIM	0.7293	±0.0368	0.8090	±0.0374	0.8248	±0.0529	0.8014	±0.0786
			PSNR	25.99	±0.47	27.86	±0.53	26.40	±1.89	27.20	±2.41
CycleGAN	median	–	SSIM	0.7975	±0.0208	0.8286	±0.0100	0.8257	±0.0202	0.8320	±0.0204
			PSNR	26.87	±0.42	28.05	±0.45	27.84	±1.14	28.27	±1.37
Dewey et al.	median	–	SSIM	0.7817	±0.0160	0.8145	±0.0101	0.8320	±0.0158	0.8346	±0.0311
			PSNR	26.51	±0.48	27.43	±0 39	27.57	±0.92	28.10	±0.89
CALAMITI	median	√	SSIM	0.8088	±0.0174†	0.8356	±0.0199	0.8369	±0.0141	0.8451	±0.0205
			PSNR	27.25	±0.47	27.69	±0.46	27.82	±0.99	27.96	±0.84
CALAMITI	fusion	–	SSIM	0.8068	±0.0161†	0.8499	±0.0184	0.8531	±0.0160†	0.8555	±0.0285†
			PSNR	27.01	±0 54	27.28	±0.43	28.02	±0.98†	28.11	±0.66
CALAMITI	fusion	√	SSIM	0.8096	±0.0222	0.8590	±0.0213	0.8599	±0.0269	0.8556	±0.0376
			PSNR	27.48	±0.49	28.48	±0.55	28.19	±1.22	28.38	±1.12

^†marks the cases that are not statistically significant based on paired Wilcoxon signed rank tests.