Table 3.

Non-i.i.d. performance for 12 fastMRI clients. The average loss values over entire validation dataset are shown for the structural similarity index measure (SSIM) and normalized root mean squared error (NRMSE). The best performance (highest SSIM, lowest NRMSE) is shown in bold font. For all sites, adaptive algorithms have the best performance.

Algorithm	Site 1		Site 2		Site 3		Site 4		Site 5		Site 6
	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE
FedAvg	0.767	0.142	0.653	0.229	0.824	0.111	0.844	0.089	0.919	0.107	0.900	0.109
FedAdam	0.768	0.142	0.677	0.183	0.821	0.117	0.845	0.093	0.928	0.096	0.901	0.104
FedYogi	0.770	0.140	0.675	0.192	0.826	0.105	0.848	0.090	0.922	0.102	0.900	0.106
FedAdaGrad	0.766	0.144	0.675	0.190	0.826	0.113	0.846	0.092	0.919	0.109	0.898	0.109
Scaffold	0.771	0.134	0.680	0.166	0.838	0.098	0.848	0.083	0.929	0.098	0.902	0.107
FL-MRCM	0.609	0.208	0.458	0.230	0.670	0.193	0.682	0.185	0.806	0.266	0.761	0.294
Centralized	0.762	0.142	0.674	0.172	0.824	0.118	0.836	0.095	0.917	0.116	0.892	0.127
Algorithm	Site 7		Site 8		Site 9		Site 10		Site 11		Site 12
	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE	SSIM	NRMSE
FedAvg	0.872	0.098	0.803	0.146	0.920	0.086	0.925	0.106	0.851	0.106	0.905	0.113
FedAdam	0.874	0.097	0.814	0.133	0.926	0.079	0.928	0.100	0.853	0.105	0.904	0.108
FedYogi	0.874	0.097	0.815	0.129	0.926	0.079	0.929	0.098	0.856	0.100	0.906	0.108
FedAdaGrad	0.873	0.097	0.811	0.137	0.924	0.082	0.929	0.105	0.854	0.103	0.903	0.111
Scaffold	0.872	0.097	0.812	0.131	0.927	0.082	0.929	0.102	0.859	0.098	0.903	0.112
FL-MRCM	0.568	0.278	0.521	0.258	0.757	0.370	0.749	0.293	0.743	0.211	0.738	0.284
Centralized	0.863	0.107	0.800	0.145	0.920	0.097	0.920	0.121	0.852	0.107	0.897	0.134