We propose an original spatio-temporal deconvolution approach for perfusion-weighted MRI applied to cerebral ischemia. The regularization of the underlying inverse problem is achieved with spatio-temporal priors and the resulting optimization problem is solved by half-quadratic minimization. Our approach offers strong convergence guarantees, including when the spatial priors are non-convex. Moreover, experiments on synthetic data and on real data collected from subjects with ischemic stroke show significant performance improvements over the standard approaches—namely, temporal deconvolution based on either truncated singular-value decomposition or ℓ 2 -regularization—in terms of various performance measures.