Brain Perfusion Analysis Method using Computed Tomography Images

Abstract

In this paper, a multi-stage perfusion calculation pipeline is suggested. It contains preprocessing algorithms for brain tissue segmentation, input artery and output vein detection, a self-supervised neural network for CT image denoising, and regularization deconvolution methods. SVD and TTV-based regularization methods were used at the last stage. The results of the comparison of these methods to classical SVD and TTV ones show that the self-supervised method outperforms others both for simulation and real data. For simulation, RMSE and SSIM metrics were used for comparison, and as for the real data, CNR metrics were compared for lesion and normal white matter areas, and for the latter ones bias and standard deviation were calculated.