Deep learning models for rapid denoising of 5D cardiac photon-counting micro-CT images

Photon-counting detectors (PCDs) are advantageous for spectral CT imaging and material decomposition because they simultaneously acquire projections at multiple energies using energy thresholds. Unfortunately, the PCD produces noisy weighted filtered backprojection (wFBP) reconstructions due to diminished photon counts in high-energy bins. Iterative reconstruction generates high quality PCD images, but requires long computation time, especially for 5D (3D + energy + time) in vivo cardiac imaging. Our recent work introduced a convolutional neural network (CNN) approach called UnetU for accurate 4D (3D + energy) photon-counting CT (PCCT) denoising at various acquisition settings. In this study, we explore how to adapt UnetU to denoise 5D in vivo cardiac PCCT reconstructions of mice. We experiment with singular value decomposition (SVD) modifications along the energy and time dimensions and replacing the U-net with a FastDVDNet architecture designed for color video denoising. All CNNs used the same group of 5D cardiac PCCT mouse sets, with 6 for training and a 7th held out for testing. All DL methods were more than 20 times faster than iterative reconstruction. UnetU Energy (which takes SVD along the energy dimension) was the most consistent at producing low root mean square error (RMSE) and spatio-temporal reduced reference entropic difference (STRRED) as well as good qualitative agreement with iterative reconstruction. This result is likely because 5D cardiac PCCT data has lower effective rank along the energy dimension than the time dimension. FastDVDNet showed promise but did not outperform UnetU Energy. Our study establishes UnetU Energy as a very accurate method for denoising 5D cardiac PCCT reconstructions that is more than 32 times faster than iterative reconstruction. This advancement enables high quality cardiac imaging with low computational burden, which is valuable for cardiovascular disease studies in mice.