Deep-learning micro-CT perfusion quantification

This study investigates the use of a convolutional neural network to perform micro-CT perfusion quantification. Perfusion CT has demonstrated substantial benefits in human medicine, being widely used for assessing cerebral blood flow in stroke patients, evaluating myocardial perfusion in cardiac diseases, and monitoring tumor vascularity in oncology. The ability to quantify perfusion metrics such as blood flow, blood volume, and mean transit time provides valuable insights into tissue viability and function, aiding diagnosis, treatment planning, and monitoring therapeutic responses. Preclinical micro-CT perfusion imaging holds significant promise in advancing our understanding of various physiological and pathological processes in small animal models. Various methods have been developed to quantify perfusion metrics from CT data, including gamma-variate parameter fittings and deconvolution methods. However, these methods have notable drawbacks, particularly their voxel-by-voxel nature, which can introduce significant noise and variability into the perfusion maps. Deep learning is a promising alternative for perfusion analysis due to its ability to learn complex patterns and relationships from large datasets. In this work, we demonstrate a deep learning approach to perfusion quantification. The network input consists of micro-CT images at 20 timepoints of time-attenuation curves. The output of the network consists of 4 parametric maps representing the numerical parameters of a gamma variate curve. The network was able to predict idealized gamma variate curves from noisy, distorted inputs with a mean absolute percent error of less than 3.4%. When applied to real data, a significant amount of noise was present as expected; however, realistic flow in the inferior vena cava and circle of Willis was visible.