Optshrink LR + S: accelerated fMRI reconstruction using non-convex optimal singular value shrinkage

Brain Informatics

Table 1 Reconstruction results with different methods on two datasets

Dataset	Method	NMSE			PSNR
Dataset	Method	6 lines	12 lines	24 lines	6 lines	12 lines	24 lines
Task-based dataset	Direct IFT	0.3088	0.2177	0.1595	3.54	6.58	9.63
	CS with wavelet sparsity [13]	0.2435	0.219	0.138	4.828	7.79	12.07
	LR + S [15]	0.1992	0.1215	0.0699	6.583	10.87	15.67
	Proposed Optshrink LR + S (r = 1)	0.0497	0.0442	0.0401	18.62	19.65	20.49
	Proposed Optshrink LR + S (r = 2)	0.0501	0.0437	0.0401	18.571	19.76	20.49
	Proposed Optshrink LR + S (r = 3)	0.0496	0.0435	0.0401	18.649	19.78	20.49
Resting-state dataset	Direct IFT	0.4067	0.286	0.1979	2.93	6.09	9.415
	CS with wavelet sparsity [13]	0.2917	0.2054	0.1193	5.42	8.465	13.19
	LR + S [15]	0.2351	0.1198	0.0576	7.321	13.16	19.52
	Proposed Optshrink LR + S (r = 1)	0.0469	0.0359	0.031	21.32	23.64	24.91
	Proposed Optshrink LR + S (r = 2)	0.0462	0.036	0.0311	21.44	23.62	24.88
	Proposed Optshrink LR + S (r = 3)	0.0473	0.0364	0.0312	21.24	23.52	24.85

Task-based data—false belief task fMRI data, subject no. 1, results on slice number 18, averaged over all time points
Resting-state data—Baltimore fMRI data, subject no. 1, results on slice number 24, averaged over all time points