Clustering-based deconvolution of brain CT perfusion data

Computed Tomography Perfusion (CTP) imaging relies on robust deconvolution methods to estimate kinetic parameters, from separation of the arterial input function and residue (i.e. tissue retention) function. While techniques based on singular value decomposition (SVD) provide computational efficiency, their lack of physiological constraints on the residue function can reduce accuracy of perfusion parameter estimation. We propose a clustering-based deconvolution framework, representing voxel-level time density curves (TDCs) as linear combinations of physiologically informed basis functions. Features extracted from TDCs guide optimal basis selection via clustering, which mitigates the high noise inherent in individual voxels, while non-negative linear least-squares and grid-search are used to estimate physiological parameters of interest. Results demonstrate improved stability and alignment with physiological expectations.