Enhancing Brain CT Perfusion Analysis Using Cluster-Based Deconvolution

Computed Tomography Perfusion (CTP) imaging plays a crucial role in the quantitative assessment of cerebral hemodynamics, diagnosis and management of stroke and other cerebrovascular diseases. Traditional deconvolution-based approaches, such as Singular Value Decomposition (SVD), offer computational efficiency, providing a rapid and direct solution for cerebral blood flow (CBF) and cerebral blood volume (CBV) estimation. However, the lack of physiological constraints on the residue function affects parameter estimation accuracy—especially under low signal-to-noise ratio conditions and in the presence of larger delays in certain tissue. To address these limitations, parametric kinetic models have been developed to provide a more physiologically meaningful representation of tracer kinetics. Techniques such as the Adiabatic Approximation to Tissue Homogeneity (AATH) model and other similar alternatives incorporate physiological constraints on the residue function, leading to improved parameter estimation. To mitigate the impact of voxel-level noise on perfusion parameter estimation, we introduce a cluster-based deconvolution framework. This approach groups spatially and kinetically similar voxels, leveraging shared hemodynamic patterns to improve stability and reduce noise sensitivity. Feature extraction from time density curves (TDCs) guides optimal cluster selection, and non-negative linear least-squares combined with grid-search optimization is used to compute and optimize perfusion parameters. In this study, we employ a brain phantom-based simulation framework to systematically evaluate the accuracy and robustness of different CTP quantification methods. The phantom allows controlled testing under various perfusion conditions, including different levels of noise and different tissue types. We compare a reference SVD-based deconvolution approach to the proposed clustering-based deconvolution framework. Simulation results demonstrate that the latter is competitive in terms of robustness and noise reduction performance, whilst maintaining computational efficiency, and offers a balance between physiological interpretability and computational feasibility. Our findings suggest that integrating clustering with deconvolution techniques could enhance the accuracy of CTP analysis, particularly in images with higher noise levels. Future work will focus on refining clustering strategies and extending validation to clinical datasets.