Arterial spin labeling MRI denoising via locally adaptive regularization with structure-guided collaborative data...

Arterial spin labeled (ASL) perfusion MRI is the only non-invasive and non-radioactive technique for measuring regional tissue perfusion. Perfusion signal in ASL MRI is derived from the difference between the spin labeled image and the spin untagged control image. Limited by the T1 decay of arterial blood, ASL MRI has an intrinsic low signal-to-noise-ratio. Solving this problem is challenging because the ground truth is often unknown, and it is difficult to preserve textures when suppressing heavy noise. In this paper we propose an unsupervised Locally Adaptive regularization with Collaborative data Selection (LACS) scheme, which exploits the high affinity between the paired label and control (L/C) images to select highly correlated contents to form the low-rank matrices. The low-rank regularization applied to such matrices could be better adapted to local structures compared with slice-level global models and more robust against noise compared with voxel-level local models. Further, we used the log-determinant of covariant matrices as the non-convex surrogate of the low-rank penalty instead of the widely used convex surrogates. We demonstrated that the adopted surrogate essentially exploits near-optimal sparsity in the underlying principal component analysis (PCA) domain without explicit training. Apparently, LACS does not rely on any ground-truth training data. When tested on a real-world ASL MRI dataset, LACS significantly improved the quality of ASL perfusion maps using just one pair of L/C images, compared with the standard pipeline that requires multiple L/C pairs. The proposed scheme could set a new benchmark for ASL MRI denoising.