Computationally efficient deep neural network for computed tomography image reconstruction

Deep neural network-based image reconstruction has demonstrated promising performance in medical imaging for undersampled and low-dose scenarios. However, it requires large amount of memory and extensive time for the training. It is especially challenging to train the reconstruction networks for three-dimensional computed tomography (CT) because of the high resolution of CT images. The purpose of this work is to reduce the memory and time consumption of the training of the reconstruction networks for CT to make it practical for current hardware, while maintaining the quality of the reconstructed images.

We unrolled the proximal gradient descent algorithm for iterative image reconstruction to finite iterations and replaced the terms related to the penalty function with trainable convolutional neural networks (CNN). The network was trained greedily iteration by iteration in the image domain on patches, which requires reasonable amount of memory and time on mainstream graphics processing unit (GPU). To overcome the local-minimum problem caused by greedy learning, we used deep UNet as the CNN and incorporated separable quadratic surrogate with ordered subsets for data fidelity, so that the solution could escape from easy local minimums and achieve better image quality.

The proposed method achieved comparable image quality with state-of-the-art neural network for CT image reconstruction on two-dimensional (2D) sparse-view and limited-angle problems on the low-dose CT challenge dataset. The difference in root-mean-square-error (RMSE) and structural similarity index (SSIM) was within [-0.23,0.47] HU and [0,0.001], respectively, with 95% confidence level. For three-dimensional (3D) image reconstruction with ordinary-size CT volume, the proposed method only needed 2 GB graphics processing unit (GPU) memory and 0.45 s per training iteration as minimum requirement, whereas existing methods may require 417 GB and 31 min. The proposed method achieved improved performance compared to total variation- and dictionary learning-based iterative reconstruction for both 2D and 3D problems.

We proposed a training-time computationally efficient neural network for CT image reconstruction. The proposed method achieved comparable image quality with state-of-the-art neural network for CT reconstruction, with significantly reduced memory and time requirement during training. The proposed method is applicable to 3D image reconstruction problems such as cone-beam CT and tomosynthesis on mainstream GPUs.