Table 1.

Quantitative performance of different methods on test DCE-MRI volumes

	Method	SSIM			NRMSE (%)
		Ktrans	ve	vp	Ktrans	ve	vp
Whole Range	DMF	0.9875 ± 0.0052	0.9960 ± 0.0010	0.9880 ± 0.0046	1.57 ± 0.43	1.41 ± 0.27	1.30 ± 0.38
	LSTM1	0.9853 ± 0.0068	0.9922 ± 0.0027	0.9806 ± 0.0159	1.47 ± 0.36	1.44 ± 0.25	1.15 ± 0.24
	LSTM2	0.9840 ± 0.0078	0.9931 ± 0.0025	0.9850 ± 0.0060	1.39 ± 0.41	1.19 ± 0.14	0.97 ± 0.18
	LSTM3	0.9850 ± 0.0070	0.9962 ± 0.0010	0.9841 ± 0.0078	1.36 ± 0.37	1.14 ± 0.17	0.97 ± 0.22
	CNN1	0.9547 ± 0.0202	0.9534 ± 0.0180	0.8946 ± 0.0267	3.07 ± 1.32	4.04 ± 1.57	3.54 ± 1.37
	CNN2	0.9629 ± 0.0159	0.9594 ± 0.0146	0.9323 ± 0.0237	2.78 ± 0.99	4.17 ± 1.32	3.16 ± 1.44
	CNN3	0.9627 ± 0.0189	0.9556 ± 0.0192	0.9302 ± 0.0352	2.76 ± 1.10	3.95 ± 1.30	2.85 ± 0.94
p	LSTM3 vs DMF	<0.05	0.54	<0.005	<0.05	<0.005	<0.005
p	LSTM3 vs CNN3	<0.005	<0.005	<0.005	<0.005	<0.005	<0.005
GTV	DMF	0.9994 ± 0.0006	0.9997 ± 0.0004	0.9994 ± 0.0008	3.85 ± 2.64	0.85 ± 0.35	0.74 ± 0.40
	LSTM3	0.9993 ± 0.0006	0.9998 ± 0.0003	0.9994 ± 0.0007	4.45 ± 2.64	0.85 ± 0.21	0.71 ± 0.24
p		0.38	<0.05	0.63	<0.05	1.00	0.45

The LSTM models were trained and tested using synthetic data with original temporal-sampling. The SSIM and NRMSE% (mean + std) with respect to the ground truth parameter maps were obtained in the whole field of view. The bold numbers indicate significant differences (p<0.05) between LSTM and DMF. GTV: gross tumor volume.