Convolutional regularization methods for 4D, x-ray CT reconstruction

Abstract

Deep learning methods have shown great promise in tackling challenging medical imaging tasks. Within the field of x-ray CT, deep learning for image denoising is of interest because of the fundamental link between ionizing radiation dose and diagnostic image quality, the limited availability of clinical projection data, and the computational expense of iterative reconstruction methods. Here, we work with 3D, temporal CT data (4D, cardiac CT), where redundancies in spatial sampling necessitate careful control of imaging dose. Specifically, using custom extensions to the Tensorflow and Keras machine learning packages, we construct and train a 4D, convolutional neural network (CNN) to denoise helical, cardiac CT data acquired in a mouse model of atherosclerosis. With the objective of accelerating iterative reconstruction, we train the CNN to map undersampled algebraic reconstructions of the 4D data to fully-sampled and regularized iterative reconstructions under mean-squared-error, perceptual loss, and low rank cost terms. Using phantom data for quantitative validation, we verify that the CNN robustly denoises static potions of the image without compromising temporal fidelity and that the CNN performs similarly to regularized, iterative reconstruction with the split Bregman method (CNN temporal RMSE: 142 HU; iterative temporal RMSE: 136 HU). Using in vivo validation and testing data excluded from CNN training, we verify that the CNN generalizes well, approximately reproducing the noise power spectrum of the iteratively reconstructed data (noise std. in water vial near heart, CNN: 62-73 HU, depending on cardiac phase; iterative: 94-100 HU), without degradation of spatial resolution (axial MTF, 10% cutoff, CNN: 2.69 lp/mm; iterative: 2.63 lp/mm). Overall, the results presented in this work represent a positive step toward realizing the promises of deep learning methods in medical imaging.