Low-Dose Dynamic CT Perfusion Denoising Without Training Data

Abstract

Cerebral perfusion maps derived from low-dose computed tomography (CT) data typically suffer from low signal-to-noise ratio (SNR). Obtaining denoised perfusion maps is critical for improved quantitative accuracy and clinical decision making. Several prior works focus on denoising the low-dose CT (LD-CT) images followed by perfusion map generation via regularized deconvolution. Recently, supervised deep neural networks (DNN) have been employed for learning a mapping between the perfusion maps obtained at low-dose and corresponding maps obtained at standard dose. Supervised learning-based methods rely on large amount of training data for improved accuracy. However, they suffer from changes in acquisition protocol and risk missing patient specific information, which is critical for applications such as stroke imaging. In this work, we address the problem of handling changes in signal characteristics while retaining patient-specific information. For this purpose, we explore a combination of self-supervised and unsupervised deep neural networks to obtain denoised perfusion maps from low-dose noisy CT data. We propose a two-stage sequential approach for obtaining improved perfusion maps to leverage statistical independence of noise in the measurement space. First, we denoise the low-dose CT projection space data using a self-supervised method. Subsequently, we reconstruct the CT images from the denoised sinograms using fast reconstruction methods such as filtered backprojection. Second, we use the improved CT images as anatomical prior to refine the noisy CBF maps obtained directly from the LD-CT data. Through empirical experiments, we show that our model is robust both qualitatively and quantitatively to changes in signal characteristics of the acquired dynamic CT data.