Investigation of Different Compressed Sensing Approaches for Respiratory Gating in Small Animal CT

Abstract

Respiratory gating is necessary in cardio-thoracic small-animal imaging because of the physiological motions that are present during scanning. When applying a low-dose protocol, fewer than 180 noisy projections may be left for the reconstruction of each respiratory phase, leading to streaking artifacts. The Prior Image Constrained Compressed Sensing (PICCS) algorithm enables accurate reconstruction of highly undersampled data when a prior image is available. The gradient domain is the preferred choice for enforcing the sparsity of the variation of each phase with respect to the prior. We evaluate three CS algorithms based on the Split-Bregman approach, with different transformations of the prior penalty function: Gradient (TV-PICCS), L1-norm (L1-PICCS), and Wavelet Transform (WT-PICCS), on low-dose data acquired on a micro-CT scanner. High-dose projection data, acquired with a cone-beam micro-CT scanner, were arranged into four gates using a software-based retrospective gating and reconstructed with an FDK algorithm. To simulate the low-dose case, we took 120 projections from each gate and added Poisson noise. Prior image was obtained by the addition of all low-dose gates with Gaussian filtering. All CS methods performed very similarly in terms of noise and resolution, greatly improving FBP (79 % noise reduction) and eliminating streaks. Pixel and wavelet domains were found to be sparser than the commonly used gradient domain. In conclusion, compressed sensing using a Split-Bregman methodology is a feasible approach to reduce dose in CT respiratory gating. The selection of the sparsity transform for the prior term does not affect resolution and noise performance but it has an influence on the final image texture: Wavelet transform showed a more natural pattern than TV and L1-norm.