Deep learning based denoising and artifact reduction in cardiac CT cine imaging
A method for cardiac computed tomography imaging acquires 4D image volumes using cardiac CT, generates synthetic low noise data from the 4D image volumes for a complete cardiac cycle by exploiting redundancies in the data, uses the generated synthetic low noise data to train a 3D U-Net with 2D+time architecture to produce denoised cardiac CT output from noisy cardiac CT input, and uses the trained 3D U-Net for deep learning based denoising of functional cardiac CT data.
1. A method for cardiac computed tomography imaging, the method comprising:
a) acquiring 4D image volumes using cardiac CT;
b) generating synthetic low noise data from the 4D image volumes for a complete cardiac cycle by exploiting redundancies in the data;
c) using the generated synthetic low noise data to train a 3D U-Net to produce denoised cardiac CT output from noisy cardiac CT input, wherein the 3D U-Net has a 2D+time architecture;
d) using the trained 3D U-Net for deep learning based denoising of functional cardiac CT data;
wherein generating synthetic low noise data comprises performing image registration within the 4D image volume by
i) selecting a lowest noise time point;
ii) performing co-registration of the lowest noise time point with other time points using non-rigid constrained co-registration.