Performance of compressed sensing for fluorine-19 magnetic resonance imaging at low signal-to-noise ratio conditions.

PURPOSE

To examine the performance of compressed sensing (CS) in reconstructing low signal-to-noise ratio (SNR) F MR signals that are close to the detection threshold and originate from small signal sources with no a priori known location.

METHODS

Regularization strength was adjusted automatically based on noise level. As performance metrics, root-mean-square deviations, true positive rates (TPRs), and false discovery rates were computed. CS and conventional reconstructions were compared at equal measurement time and evaluated in relation to high-SNR reference data. F MR data were generated from a purpose-built phantom and benchmarked against simulations, as well as from the experimental autoimmune encephalomyelitis mouse model. We quantified the signal intensity bias and introduced an intensity calibration for in vivo data using high-SNR ex vivo data.

RESULTS

Low-SNR F MR data could be reliably reconstructed. Detection sensitivity was consistently improved and data fidelity was preserved for undersampling and averaging factors of α = 2 or = 3. Higher α led to signal blurring in the mouse model. The improved TPRs at α = 3 were comparable to a 2.5-fold increase in measurement time. Whereas CS resulted in a downward bias of the F MR signal, Fourier reconstructions resulted in an unexpected upward bias of similar magnitude. The calibration corrected signal-intensity deviations for all reconstructions.

CONCLUSION

CS is advantageous whenever image features are close to the detection threshold. It is a powerful tool, even for low-SNR data with sparsely distributed F signals, to improve spatial and temporal resolution in F MR applications.